- Cisco Talos's code analysis of a Windows implant, CLOSEDQUORUM, shows it asking up to four commercial AI models what to do next and running the action with the most votes, with no attacker-run server.
- No one has seen the full sequence run, and the public build is an inert template with placeholder credentials, so Talos treats it as an early, limited example.
- Talos advises detecting combinations of behavior, not domains. Ask your security team whether it can flag one host calling several AI services alongside credential theft or new persistence.
Cisco Talos's static analysis of a Windows implant it calls CLOSEDQUORUM shows a design that, once deployed, asks up to four commercial AI models what to do next, tallies their votes and executes the winner. It needs no further commands from a human operator and no dedicated attacker-run C2 server. There is a limit on what the findings show: they come from reading the code, and no one has watched the whole sequence run from start to finish. The publicly distributed build is an inert template with placeholder credentials. Talos released the analysis on September 22, 2026, as the first in a series. It accompanies CAIRN, an open-source research toolkit the team built to track malware that incorporates AI. To Talos's knowledge, CLOSEDQUORUM is the earliest Windows implant of this kind to be documented publicly.
No single server to block
Traditional command and control rests on infrastructure the attacker runs: a domain, an IP address, a listener. Talos notes that defenders track those domains, threat feeds publish those IPs, and certificate transparency logs can expose new infrastructure before it is used. CLOSEDQUORUM has no single server of that kind to find. It calls the public endpoints of DeepSeek, Qwen, Mistral and Google Gemini, which Talos says thousands of legitimate applications use daily. Each provider answers in a fixed JSON format, and the action with the most votes wins. A tie is settled in a set order: DeepSeek, then Qwen, Mistral and Gemini.
Network indicators do not disappear. Discord carries the implant's real-time reports and its stolen data. But Talos's guidance is to build detection around behavior rather than domain blocking, because legitimate applications may reach the same services on their own.
What the implant is built to take
The Go-compiled binary is aimed at credentials and cryptocurrency. Talos lists LSASS memory capture, login data from Chrome, Edge and Firefox, and wallet files for MetaMask, Exodus and Ethereum among its techniques. The models can only choose from capabilities the developer has already coded, and a reply that does not map to one is discarded. Persistence combines a registry run key, a scheduled task and a WMI event subscription that fires every 60 seconds and runs a PowerShell script from the Windows temp folder. Stolen data is encrypted, base64-encoded and posted to an operator's Discord webhook in segments.
A limited but concrete precedent
Talos is explicit about the evidence. Its code analysis confirms the full decision loop, and development builds show provider credentials being injected at build time. What no one has seen is the loop running start to finish: the version circulating publicly ships with placeholder credentials and does nothing useful as distributed.
Talos also says the implant appears to work as an operator-configured service, with the developer producing a customized binary and a separate exfiltration channel for each operator. If that model holds, a buyer deploys the binary and need not stay online while the model panel runs the attack. That is the sense in which the human is removed: the operator still acquires, configures and deploys the tool, but no longer issues commands afterward. Talos frames CLOSEDQUORUM as an early, limited example. It reads the implant as a demonstration rather than a sophisticated tool: handing a bounded phase of an intrusion to models is achievable with ordinary API access. It also lists weaknesses: provider refusals, rate limits, malformed output and predictable tie-breaking.
What to ask your security team
Talos points defenders toward combinations of behavior, not single indicators. That suggests questions for the team rather than a purchase order. Can our monitoring flag one host contacting several external LLM APIs, especially alongside LSASS access, injection into suspended processes or new WMI persistence? Can we see Discord webhook traffic from processes that also touch LSASS or create WMI persistence? Can our detection tell approved AI-tool usage apart from an unapproved process reaching the same provider endpoints? And does our incident response plan account for malware that keeps acting after the person who deployed it has gone offline?
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Cisco Talos.





