OpenAI has banned a cluster of accounts it assesses were operated by a Russian-speaking cybercriminal group that turned ChatGPT into an unlikely development partner for crafting stealthy malware. The activity, detailed in an October 2025 report, reveals a persistent effort to use the AI to prototype and troubleshoot malicious tools, even though the model’s safety guardrails directly blocked overtly criminal requests.
Instead, the operator played a clever game of legos with malicious code. Unable to get the AI to write a finished remote-access trojan (RAT) or keylogger, the threat actor broke the process into smaller, legitimate-seeming building blocks. They prompted the model to convert executables into shellcode, design in-memory loaders, and generate scripts to parse browser credentials. These components, benign in isolation, were then assembled off-platform into a functional toolkit for credential theft and crypto asset raiding.
The sophistication was uneven but telling. Some prompts required deep Windows internals knowledge—leveraging the model for DPAPI and AES-GCM cookie decryption or Chrome DevTools protocol automation. Others were remarkably blunt, such as generating code for “crypter” patterns that insert junk sequences to evade antivirus detection. The operator iterated on the same code across a small number of accounts, a pattern consistent with a focused development cycle rather than casual experimentation. OpenAI shared indicators with industry partners but noted the capabilities were not novel; the models refused direct exploit requests, and the resulting techniques are already widely documented in public forums.
