Cisco Talos has released an open-source framework for identifying malware that incorporates AI services. Called CAIRN, it looks for technical traces left by model integration, assigns samples comparable identifiers and groups them by shared traits so defenders can track how the technique is spreading.
The framework has already surfaced CLOSEDQUORUM, Windows malware that asks as many as four language models—DeepSeek, Qwen, Mistral and Google Gemini—what to do next. It combines their responses into a consensus and has no mechanism for a human operator to intervene. Redundant model access also lets it continue if one service is unavailable.
Researchers found links to criminal forums discussing credit-card fraud, and the software is designed to steal credentials and cryptocurrency. They could not identify its developer or confirm that it has been used in a real attack. That uncertainty is important: autonomous design does not prove operational success.
Cisco researcher Ryan Fetterman found about 20 additional AI-integrated malware examples while building CAIRN, beyond the small number previously documented. He still describes the field as experimental, but the framework gives defenders a shared way to classify early samples before these methods become routine.