MQsTTang: Mustang Panda’s latest backdoor treads new ground with Qt and MQTT

ESET researchers have analyzed MQsTTang, a new custom backdoor that we attribute to the Mustang Panda APT group. This backdoor is part of an ongoing campaign that we can trace back to early January 2023. Unlike most of the group’s malware, MQsTTang doesn’t seem to be based on existing families or publicly available projects.

Mustang Panda is known for its customized Korplug variants (also dubbed PlugX) and elaborate loading chains. In a departure from the group’s usual tactics, MQsTTang has only a single stage and doesn’t use any obfuscation techniques.

Victimology

We have seen unknown entities in Bulgaria and Australia in our telemetry. We also have information indicating that this campaign is targeting a governmental institution in Taiwan. However, due to the nature of the decoy filenames used, we believe that political and governmental organizations in Europe and Asia are also being targeted. This would also be in line with the targeting of the group’s other recent campaigns. As documented by fellow researchers at Proofpoint, Mustang Panda has been known to target European governmental entities since at least 2020 and has increased its activity in Europe even further, since Russia’s invasion of Ukraine. Figure 1 shows our view of the targeting for this campaign.

Figure 1. Map showing known and suspected targets of MQsTTang

Attribution

We attribute this new backdoor and the campaign to Mustang Panda with high confidence based on the following indicators.

We found archives containing samples of MQsTTang in two GitHub repositories belonging to the user YanNaingOo0072022. Another GitHub repository of the same user was used in a previous Mustang Panda campaign described by Avast in a December 2022 blogpost.

One of the servers used in the current campaign was running a publicly accessible anonymous FTP server that seems to be used to stage tools and payloads. In the /pub/god directory of this server there are multiple Korplug loaders, archives, and tools that were used in previous Mustang Panda campaigns. This is the same directory that was used by the stager described in the aforementioned Avast blogpost. This server also had a /pub/gd directory, which was another path used in that campaign.

Some of the infrastructure used in this campaign also matches the network fingerprint of previously known Mustang Panda servers.

Technical analysis

MQsTTang is a barebones backdoor that allows the attacker to execute arbitrary commands on a victim’s machine and get the output. Even so, it does present some interesting characteristics. Chief among these is its use of the MQTT protocol for C&C communication. MQTT is typically used for communication between IoT devices and controllers, and the protocol hasn’t been used in many publicly documented malware families. One such example is Chrysaor, also known as Pegasus for Android. From an attacker’s perspective, one of MQTT’s benefits is that it hides the rest of their infrastructure behind a broker. Thus, the compromised machine never communicates directly with the C&C server. As seen in Figure 2, this capability is achieved by using the open source QMQTT library. This library depends on the Qt framework, a large part of which is statically linked in the malware. Using the Qt framework for malware development is also fairly uncommon. Lazarus’s MagicRAT is one of the rare recently documented examples.

Figure 2. RTTI showing classes from the QMQTT library

MQsTTang is distributed in RAR archives which only contain a single executable. These executables usually have names related to Diplomacy and passports such as:

• CVs Amb Officer PASSPORT Ministry Of Foreign Affairs.exe
• Documents members of delegation diplomatic from Germany.Exe
• PDF_Passport and CVs of diplomatic members from Tokyo of JAPAN.eXE
• Note No.18-NG-23 from Embassy of Japan.exe

These archives are hosted on a web server with no associated domain name. This fact, along with the filenames, leads us to believe that the malware is spread via spearphishing.

So far, we have only observed a few samples. Besides variations in some constants and hardcoded strings, the samples are remarkably similar. The only notable change is the addition of some anti-analysis techniques in the latest versions. The first of these consists of using the CreateToolhelp32Snapshot Windows API function to iterate through running processes and look for the following known debuggers and monitoring tools.

• cheatengine-x86_64.exe
• ollydbg.exe
• ida.exe
• ida64.exe
• radare2.exe
• x64dbg.exe
• procmon.exe
• procmon64.exe
• procexp.exe
• processhacker.exe
• pestudio.exe
• systracerx32.exe
• fiddler.exe
• tcpview.exe

Note that, while the malware is a 32-bit executable, it only checks for the presence of x64dbg and not its 32-bit counterpart, x32dbg.

The second technique uses the FindWindowW Windows API to look for the following Window Classes and Titles used by known analysis tools:

• PROCMON_WINDOW_CLASS
• OLLYDBG
• WinDbgFrameClass
• OllyDbg - [CPU]
• Immunity Debugger - [CPU]

When executed directly, the malware will launch a copy of itself with 1 as a command line argument. This is repeated by the new process, with the argument being incremented by 1 on every run. When this argument hits specific values, certain tasks will be executed. Note that the exact values vary between samples; the ones mentioned below correspond to the sample with SHA-1 02D95E0C369B08248BFFAAC8607BBA119D83B95B. However, the tasks themselves and the order in which they are executed is constant.

Figure 3 shows an overview of this behavior along with the tasks that are executed when the malware is first run.

Figure 3. Execution graph showing the subprocesses and executed tasks

Table 1 contains a list of the tasks and the value at which each of them is executed. We will describe them in further detail in the upcoming paragraphs.

Table 1. Tasks executed by the backdoor

If any analysis tool or debugger is detected using the techniques we described previously, the behavior of task 1 is altered and tasks 2, 3, and 4 are skipped entirely.

Task 1: C&C communication

As was previously mentioned, MQsTTang communicates with its C&C server over the MQTT protocol. All observed samples use 3.228.54.173 as broker. This server is a public broker operated by EMQX, who also happen to be the maintainers of the QMQTT library. This could be a way to make the network traffic seem legitimate and to hide Mustang Panda’s own infrastructure. Using this public broker also provides resiliency; the service is unlikely to be taken down because of its many legitimate users and, even if the current C&C servers are banned or taken down, Mustang Panda could spin up new ones and use the same MQTT topics without disrupting MQsTTang’s operation.

However, this campaign could also be a test case by Mustang Panda before deciding whether to invest the time and resources to set up their own broker. This is supported by the low number of samples we’ve observed and the very simple nature of MQsTTang.

As shown in Figure 4, the malware and C&C server use two MQTT topics for their communication. The first one, iot/server2, is used for communication from the client to the server. The second one is used for communication from the server to the client. It follows the format iot/v2/<Unique ID> where <Unique ID> is generated by taking the last 8 bytes, in hex form, of a UUID. If any analysis tool is detected, server2 and v2 are respectively replaced with server0 and v0. This is likely in order to avoid tipping off defenders by entirely aborting the malware’s execution early.

Figure 4. Simplified network graph of the communication between the backdoor and C&C server

All communication between the server and the client uses the same encoding scheme. The MQTT message’s payload is a JSON object with a single attribute named msg. To generate the value of this attribute, the actual content is first base64 encoded, then XORed with the hardcoded string nasa, and base64 encoded again. We will describe the exact format of these payloads in the relevant sections.

Upon first connecting to the broker, the malware subscribes to its unique topic. Then, and every 30 seconds thereafter, the client publishes a KeepAlive message to the server’s topic. The content of this message is a JSON object with the following format:

{
    "Alive": "<malware’s uptime in minutes>",    
    "c_topic": "<client’s unique topic>"
}

When the server wants to issue a command, it publishes a message to the client’s unique topic. The plaintext content of this message is simply the command to be executed. As shown in Figure 5, the client executes the received command using QProcess::startCommand from the Qt framework. The output, obtained using QProcess::readAllStandardOutput, is then sent back in a JSON object with the following format:

{
    "c_topic": "<client’s unique topic>",
    "ret": "<Command output>"
}

Figure 5. Execution of received commands using the QProcess class

Since only the content of standard output is sent back, the server will not receive errors or warnings. From the server’s point of view, a failed command is thus indistinguishable from a command that simply produces no output unless some sort of redirection is performed.

Tasks 2 and 3: Copying the malware

The second and third tasks are fairly similar to each other. They copy the malware’s executable to a hardcoded path; c:\users\public\vdump.exe and c:\users\public\vcall.exe respectively. The filenames used are different for each sample, but they are always located in the C:\users\public directory.

In the second task, the newly created copy is then launched with the command line argument 97.

Task 4: Establishing persistence

Persistence is established by the fourth task, which creates a new value qvlc set to c:\users\public\vcall.exe under the HKCU\Software\Microsoft\Windows\CurrentVersion\Run registry key. This will cause the malware to be executed on startup.

When MQsTTang is executed on startup as c:\users\public\vcall.exe, only the C&C communication task is executed.

Conclusion

The Mustang Panda campaign described in this article is ongoing as of this writing. The victimology is unclear, but the decoy filenames are in line with the group’s other campaigns that target European political entities.

This new MQsTTang backdoor provides a kind of remote shell without any of the bells and whistles associated with the group’s other malware families. However, it shows that Mustang Panda is exploring new technology stacks for its tools. It remains to be seen whether this backdoor will become a recurring part of the group’s arsenal, but it is one more example of the group’s fast development and deployment cycle.

IoCs

Files

Network

Github repositories

• https://raw.githubusercontent[.]com/YanNaingOo0072022/14/main/Documents.rar
• https://raw.githubusercontent[.]com/YanNaingOo0072022/ee/main/CVs Amb.rar

MITRE ATT&CK techniques

This table was built using version 12 of the MITRE ATT&CK framework.