Cyb3rius
@cyb3rius
· Sep 24
Can malware let AI models vote on its next move?Yes. Cisco Talos disclosed on September 22 that a Windows malware family called CLOSEDQUORUM outsources decisions to as many as four large language models and then acts on their majority vote, with no attacker-run command-and-control server required.What happened: Researchers at Cisco Talos analyzed CLOSEDQUORUM, a Windows implant that sends prompts to up to four AI models and takes the winning vote as its next instruction. Rather than beaconing to a fixed C2 domain, the malware effectively turns model endpoints into its decision layer, which makes infrastructure takedowns far less useful for defenders.Key numbers: Up to 4 AI models can vote on a single malware decision1 majority vote determines the implant's next action0 traditional C2 servers are needed to task the malwareWhy it matters: If intrusions no longer depend on attacker-owned infrastructure, domain blocking and sinkholing stop working, and defenders must detect the model queries themselves.Bottom line: When malware starts taking votes, the command-and-control server becomes optional and the AI prompt becomes the new attack surface.
0
