Security researchers at Cisco Talos have identified a new Windows malware strain named CLOSEDQUORUM. Unlike previous tools that used AI only for coding or phishing, this malware integrates multiple large language models directly into its command and control structure to make autonomous decisions.
The malware interacts with four specific models: DeepSeek, Qwen, Mistral, and Google Gemini. By sending requests to these models and analyzing their feedback, the program determines its next steps. This multi-model approach provides redundancy, allowing the malware to function even if one AI service becomes inaccessible.
CLOSEDQUORUM operates without traditional remote control entry points, relying instead on AI feedback to plan its actions. Its primary objectives include stealing user credentials and cryptocurrency assets. Researchers have linked the malware to credit card fraud activities dating back to 2025, though the developers remain unidentified.
To track such threats, Cisco Talos developed an open-source framework called CAIRN. This system analyzes metadata and technical characteristics to identify unique digital fingerprints left by AI-integrated malware. It helps researchers classify new attack patterns and monitor how attackers incorporate language models into their infrastructure.
