ClosedQuorum malware uses AI voting to automate Windows attacks

Security researchers at Cisco Talos have identified a new Windows malware prototype named ClosedQuorum that utilizes a voting mechanism among multiple artificial intelligence models to determine its next operational steps. Unlike traditional malware that relies on continuous commands from a human operator, ClosedQuorum automates decision-making by querying several large language models.

The malware uses a quorum of four AI models—DeepSeek, Qwen, Mistral, and Gemini—to select actions from a predefined menu, such as collecting files, injecting code into memory, or spreading through a network. Each model receives basic host information, and the action receiving the most votes is executed. In the event of a tie, a specific hierarchy (DeepSeek, then Qwen, Mistral, and Gemini) is used to resolve the deadlock.

This approach aims to reduce the operational burden on human attackers and bypass traditional security measures by using legitimate AI services and Discord channels for data exfiltration rather than dedicated command-and-control servers. While the analyzed version appears to be a proof of concept with test keys disabled, it highlights a shift toward integrating AI directly into the malware execution flow to automate reconnaissance and tactical movements.