ClosedQuorum malware uses AI quorum voting for autonomous attacks

Researchers have identified ClosedQuorum, a new Windows-based malware strain that utilizes an autonomous AI-driven command-and-control (C2) framework. Discovered by Cisco Talos, the implant is believed to be the first publicly documented malware to use large language models (LLMs) to make post-compromise decisions without human intervention.

The malware employs a quorum voting system involving multiple LLM providers, including Google Gemini, DeepSeek, Qwen, and Mistral. These models act as a panel of strategists, casting votes on specific executable actions such as stealing credentials, injecting shellcode, or establishing persistence. In the event of a tie, the system follows a predefined hierarchy where DeepSeek holds precedence, followed by Qwen, Mistral, and Gemini.

ClosedQuorum has been detected on over 10,000 endpoints across the finance, healthcare, and manufacturing sectors. By using AI to select actions, the malware achieves a behavioral variability that helps it evade traditional signature-based antivirus and EDR tools, reportedly reducing the mean time to compromise by approximately 40%. To combat this threat, Cisco Talos has released the open-source CAIRN hunter tool to detect anomalous model-invocation patterns.