AI-Powered Malware Operates Without Human Commands, Worrying Security Teams

Cisco Talos researchers have discovered a Windows implant called CLOSEDQUORUM that uses a panel of four AI models to make tactical decisions without human commands. This is the first publicly documented Windows implant to use this approach for command and control.

The malware, named CLOSEDQUORUM, sends information about a compromised host to up to four commercial AI providers: DeepSeek, Qwen, Mistral, and Google Gemini. Each model returns a structured response, and CLOSEDQUORUM executes whichever action wins the plurality vote.

CLOSEDQUORUM targets Windows credentials, browser-stored passwords, and cryptocurrency wallet data. The multi-provider design creates redundancy, allowing the implant to continue querying other models if one AI service is unavailable.

Talos released an open-source toolkit called CAIRN to detect, classify, and track malware that integrates AI services. During development and testing, CAIRN uncovered approximately 20 additional AI-integrated malware samples, exceeding the nine families previously documented publicly.