OpenAI has taken action against accounts associated with Russian-speaking criminal groups, as detailed in its October 2025 report. These groups were observed using ChatGPT to construct malware loaders, evasion layers, credential-theft scripts, and command-and-control (C2) infrastructure. Separately, in its June 2025 report, OpenAI described banning Russian-language accounts under "Operation ScopeCreep" for similar activities, including developing Windows malware and debugging code. A third October 2025 report from OpenAI also identified and banned Korean-language accounts attempting to use models for malware and C2 development, debugging, phishing, and credential-theft workflows.
Key Points
- OpenAI banned ChatGPT accounts linked to Russian-speaking criminal groups for developing malware loaders, evasion layers, credential-theft scripts, and C2 infrastructure.
- The Russian-speaking threat actor used ChatGPT to prototype and troubleshoot technical components for post-exploitation and credential theft, including converting executables to shellcode and parsing browser credentials.
- OpenAI's October 2025 report noted that models refused direct requests for malicious content, leading operators to elicit building-block code.
- "Operation ScopeCreep," detailed in OpenAI's June 2025 report, involved a Russian-speaking actor using ChatGPT to develop Windows malware, debug code, and set up C2 infrastructure.
- The "ScopeCreep" actor used temporary email addresses and new ChatGPT accounts for each incremental code improvement, distributing malware via a repository impersonating a legitimate gaming tool.
- OpenAI's October 2025 report on Korean-language operators identified attempts to use models for Windows API hooking, browser credential and cookie access workflows, and drafting phishing emails.
- No evidence was found that access to OpenAI's models provided these actors with novel capabilities beyond publicly available resources, and models refused clearly malicious requests.
Context
According to OpenAI, the Russian-speaking criminal groups used ChatGPT for technical tasks such as low-level PE / Win32 API handling and Chrome DevTools automation. The Korean-language operators primarily engaged with models in Korean, focusing on specific use cases per account, such as converting Chrome extensions or configuring Windows Server VPNs. OpenAI stated that these activities often fell into a "gray zone" of dual-use capabilities, where legitimate applications like software debugging could be repurposed maliciously.
Why It Matters
These incidents highlight the ongoing challenge of preventing the misuse of AI models for malicious cyber activities. Builders and developers must consider the dual-use nature of AI capabilities and the potential for threat actors to leverage models for tasks like malware development, even when direct malicious content generation is blocked. The reports indicate that AI can accelerate development workflows for threat actors, necessitating robust detection and disruption strategies.
What To Do
- Note that OpenAI's models are designed to refuse direct requests for malicious content.
- Understand that threat actors may use models to generate building-block code that can be assembled into malicious workflows.
- Watch for patterns of iterative code development and debugging across conversations, as observed in the Russian-speaking criminal group's activities.
- Review OpenAI's October 2025 and June 2025 reports for detailed examples of how threat actors attempt to use AI for cyber operations.
Keep Exploring
/atlas/claude-family /atlas/gpt-family /techniques/ptcf /techniques/system-user-separation /techniques/constraints /studio?pack=foundation