On LinkedIn
Somewhere at Cisco Talos, someone found malware that holds a vote before it robs you.
It is called CLOSEDQUORUM. Every five to fifteen minutes it messages four commercial AI models, DeepSeek, Qwen, Mistral and Gemini, describes the machine it landed on, and asks what to do next. Steal credentials? Inject into another process? Dig in and stay? Majority wins. Answers in the wrong format get thrown out, because even criminals have standards.
It is Who Wants to Be a Millionaire, except the contestant is dumping your LSASS memory and the lifeline is “ask the audience”.
The funny part: it is loud. It phones out constantly, to services your users are already talking to all day, which means egress control and behavioural detection still have something to grab.
That is a genuine gift. Take it while it lasts.
The part that is not funny: your signature describes one run. The next run votes differently. You are no longer fingerprinting malware, you are fingerprinting a mood. And when those models move on-device, and they will, the chatter goes quiet and the gift expires.
Detection has always described yesterday’s sample. This time yesterday’s sample never existed.
So the unglamorous question is the only one left. If this lands on a laptop in your network tomorrow, how far does it get before something stops it? Not what your policy says. What you would actually bet.
The Talos write-up is worth the read: The Closed Quorum: inside the first reported autonomous AI C2 implant.
One detail that did not fit above: the vote is not even democratic. Ties go to DeepSeek. If DeepSeek is unreachable, Qwen decides, and so on down a fixed order. There is a chain of command inside the thing.
They also released the tooling they hunted it with, CAIRN.