A committee of models picks the malware's next move

The image that makes this land is a bank robbery crew that pauses before each job to take a vote on *how* to do it. This malware, which Cisco Talos named CLOSEDQUORUM, is a Windows implant that acts the same way: before it carries out a command-and-control decision, it puts the question to a committee of up to four AI models (DeepSeek, Qwen, Mistral and Gemini among the names in play, with a DeepSeek-first tie-break). The models vote; the implant acts.

It is the first publicly documented Windows binary built around that idea — an agentic decision loop running inside a malicious program, where the "brain" making the calls is a panel of off-the-shelf language models rather than a hard-coded instruction set.

The designers' apparent reasoning is a kind of robustness by committee: no single model's judgement, so a panel that has to agree looks harder to steer or jam. That is a questionable bet — a vote is only as trustworthy as the models that sit on it, and a committee with a shared blind spot still votes wrong together.

For the log, though, the point is this: the decision-making that used to be the malware author's job, coded in, is now being delegated to an external AI panel at runtime. Malware has started to *think out loud* — and it does so with a quorum.

Talos is explicit that this is a reverse-engineered binary, and that they do not have confirmation of in-the-wild deployment — nor can they confirm the developer. So this is a case study of a design as much as of a deployed threat, and it is logged as such. The invention is the part worth recording: the machinery for putting C2 to a vote now exists and has been written up.