Yoba malware uses fake AI tools to deploy cryptominers
A malware campaign named Yoba uses fake GitHub repositories for AI tools like GPT-5.6 Sol to distribute cryptominers via DLL sideloading and Telegram-driven command channels.

Check Point Research has identified a new malware campaign named Yoba that uses fake GitHub repositories promoting AI tools like GPT-5.6 Sol, GPT-6 Astra, and Claude Opus to distribute malicious payloads. The campaign, tracked since October 27, leverages counterfeit ChatGPT and Claude applications to spread malware through signed binaries and covert communication channels.
Delivery Mechanism
The malware is delivered through ZIP files hosted on GitHub Pages, containing legitimate signed executables from OpenAI or Anthropic alongside malicious DLLs that exploit DLL sideloading techniques. Each Yoba ZIP contains a legitimate signed Setup.exe from OpenAI or Anthropic alongside a malicious winmm.dll. The fake projects use GitHub Pages and download buttons linking to ZIP files in GitHub releases.
Three malicious ZIP files were released:
| File Name | Size |
|---|---|
| sol-setup.zip | 269 bytes |
| Setup-Claude.zip | 82 bytes |
| GPT-6-Astra-Setup.zip | 352 bytes |
The ZIP files lack password protection and have low detection rates due to built-in anti-analysis checks. The malware was distributed through GitHub repositories with GitHub Pages landing pages, some ranking in Google's top search results due to SEO poisoning. QiAnXin reported similar fake DeepSeek TUI repositories hiding malware in 7z archives and counterfeit installers for Claude, Grok, and WormGPT. An earlier fake Claude website distributed a ZIP called Claude-Pro Relay, whose installer dropped a backdoor named Beagle.
Technical Execution
The dropper YobaDropper uses DLL sideloading to load malicious components into trusted processes, checks system language and privileges, and establishes persistence via scheduled tasks disguised as Microsoft components. The malware chain bundled legitimate signed OpenAI and Anthropic executables with a malicious DLL. Check Point said: "The fake applications bundled legitimate signed OpenAI and Anthropic executables with a malicious dynamic link library (DLL)."
YobaDropper used DLL sideloading to load malicious components, checked host process name and administrator privileges, and verified keyboard layouts to avoid CIS countries except Ukraine. The dropper employed lengthy sleeps to hinder automated analysis and showed indicators suggesting Russian origins. It created a ProgramData directory, copied a legitimate Windows executable, and established persistence via Scheduled Tasks running under the Local System account, disguised as a Microsoft task. The dropper stalls for about 12 days unless launched by Setup.exe and skips systems using Russian, Belarusian, Kazakh, and other CIS-region keyboard layouts. A scheduled task labeled as a Microsoft component is created to run three minutes after boot.
The anti-sandbox technique checks if the process name is Setup.exe using DJB2 hashing with seed 0x1505. The malware checks for CIS country location via keyboard layout, excluding Ukrainian (ID 0x22) from the CIS list. It checks if the process is running as Administrator, displaying 'Please run this program as Administrator.' if not. Malicious releases were uploaded on August 30, 31, and September 4 based on commit data.
Command and Control
YobaLoader retrieves encrypted commands via Telegram images and Google DNS-over-HTTPS, decrypts them using AES and Ed25519 signatures, and uses domain generation algorithms to fetch payloads from Catbox. The second-stage YobaLoader used Telegram and DNS-over-HTTPS (DoH) to fetch payloads without dedicated C2 infrastructure.
YobaLoader extracted data from a Telegram image to generate domains via a domain generation algorithm and queried Google's DoH service for encrypted DNS data, which was cryptographically validated before loading the next-stage payload. The DGA algorithm generates domains in the format '$adjective-$noun-$verb-$digit.dedyn[.]io' using three word lists and a Telegram channel image as seed. The loader queries Google’s DNS-over-HTTPS service for a hidden command in a DNS TXT record, appearing as normal encrypted traffic.
The downloader uses the Telegram channel 't.me/yoba_fanclub' and parses the profile image URL to generate the DGA seed. It queries Google's DNS-over-HTTPS API for a command via DNS TXT record, decrypts it with AES-ECB using key '1d0d1368b4d53f3a6322cf40e84e63db', and validates with Ed25519 signature using public key 'b5c2d20623d4e6647d06a42864edd8629f82c5c384b243fdaeeeca24b3873d64'. The decrypted response contains the URL 'https://files.catbox[.]moe/z5zpy9.dat'. Commands are verified with an Ed25519 signature and decrypted with AES before being executed.
Impact and Response
The final payload is an XMRig cryptominer, with future waves potentially delivering ransomware or data theft tools, prompting Check Point to urge caution when installing AI applications from unverified sources. The payload delivered was an XMRig cryptocurrency miner connecting via Tor, though YobaLoader is generic and could deliver other executables or PowerShell payloads in future waves.
Yash Kadakia, founder and CTO at Security Brigade, stated that valid digital signatures only confirm executable authenticity and integrity, not the trustworthiness of loaded components or distribution packages. He emphasized that a signed executable is not the same as a trusted execution chain. Enterprises should make approved AI tools easier to obtain to prevent shadow IT, review permanent local administrator privileges on developer workstations, extend controls beyond executable signatures to DLL loading and execution behavior, and monitor software from unverified repositories or user-writable directories. The campaign is particularly relevant to developer environments due to risks to source code, API keys, cloud credentials, and internal infrastructure access.
Check Point advised caution when installing AI applications from unverified repositories and recommended restricting execution of software from unknown sources. MITRE explains DLL sideloading hides malicious actions under legitimate processes, and signed software is often trusted by security tools, making detection difficult. A prior Yoba variant was observed in June using a fake Cursor AI coding-agent application. Future waves of the Yoba campaign may deliver malware families beyond cryptocurrency miners, such as ransomware or data-theft tools.





