Executive Summary
ClosedQuorum represents a significant shift in malware architecture by delegating tactical command-and-control (C2) decisions to a panel of commercial Large Language Models (LLMs), including DeepSeek, Qwen, Mistral, and Google Gemini. Discovered by Cisco Talos, this Windows-based implant uses a plurality voting system among these AI models to decide whether to steal credentials, establish persistence, or inject shellcode.
Technically, the malware functions by gathering host reconnaissance data and passing it as context to the AI models via structured JSON prompts. By leveraging legitimate AI provider APIs for C2, the malware avoids traditional indicator-based detection like blacklisted domains or IPs. Tie-breaking logic is deterministic, prioritizing DeepSeek, then Qwen, then Mistral, then Gemini.
While currently observed in a 'credentials-as-a-service' model with potential links to criminal carding forums, the primary threat is the displacement of human effort. This enables attack chains to scale and operate continuously without a human-in-the-loop, potentially bypassing traditional sandbox and detection window expectations through randomized polling and autonomous reasoning.