Cisco Talos (opens in new tab) on Tuesday published research on a Windows malware sample it calls CLOSEDQUORUM, describing an implant designed so that after deployment it can ask commercial large language models what to do next rather than taking live commands from an attacker-run command-and-control server.
According to Talos, the binary can query up to four providers—DeepSeek, Qwen, Mistral, and Google Gemini—tally their structured replies, and carry out the plurality vote among a fixed set of actions labeled steal, inject, persist, and move. WIRED (opens in new tab) and The Hacker News (opens in new tab), summarizing the same research, said the design aims to steal Windows credentials, browser passwords, and cryptocurrency wallet data, with decisions and loot reported through a Discord webhook.
Talos said CLOSEDQUORUM is, to its knowledge, the first publicly documented Windows implant to apply this multi-model voting approach to tactical C2. That “first” is the researchers’ claim, echoed by secondary coverage; The Event Record has not independently verified priority. Talos also said it has not confirmed in-the-wild deployment. The publicly observed distribution build uses placeholder API keys and a dummy webhook, and Talos said it did not observe a complete end-to-end run of the architecture.
The firm said it found the sample with CAIRN, an open-source toolkit it released the same day to track AI-integrated malware. Artifacts in the binary, Talos wrote, linked the developer to criminal-forum postings about carding dating to 2025. Delivery methods were not described. Defenders, Talos argued, should watch for behavioral combinations—unexpected multi-provider AI traffic plus LSASS access, injection, or WMI persistence—rather than simply blocking AI vendor domains.