Rather than receiving orders from an attacker-operated server, the malware is built to consult the models, which then collectively decide the next action. The choices available to those models reportedly include stealing Windows credentials, saved browser passwords, and crypto wallet data.

Several important caveats apply to the findings. Talos has not observed this model-voting setup work from start to finish, and the publicly available version of the malware does not function as-is. In other words, the code as it exists in the wild falls short of the fully operational design the researchers describe.

Key points:

• CLOSEDQUORUM is a Windows malware family that uses AI models to decide its behavior.
• Up to four AI models can participate in the vote.
• Potential actions include credential theft, browser password theft, and crypto wallet data theft.
• Cisco Talos has not seen the full chain execute successfully.
• The public version of the malware does not work in its current state.

At this stage, there are no specific indicators of compromise, affected version lists, or mitigation guidance beyond what the research describes, so defenders should treat the report as an early look at an emerging technique rather than a confirmed, fully functional threat.

Source: The Hacker News

Would you trust an AI model to make the final call on your organization's detection and response decisions?