OpenAI has banned accounts suspected of being linked to Russian-speaking criminal groups that were using its AI models to develop and refine malware.

The actors leveraged the AI to build components for remote-access trojans, credential stealers, and evasion techniques. While the AI refused direct requests for malicious content, the users elicited building-block code for tasks like converting executables to shellcode, designing in-memory loaders, and parsing browser credentials. They also used the AI for obfuscation, crypter patterns, and exfiltration helpers, such as Telegram bots.

The threat actor focused on technical tasks, including Win32 API functions and browser cookie handling, iterating on code across multiple conversations. OpenAI disabled the associated accounts and shared indicators with partners, noting that the AI did not provide novel capabilities beyond publicly available resources. The models refused clearly malicious requests.