Overview
- Cisco Talos released the CAIRN framework and publicly disclosed findings Tuesday showing a Windows implant named CLOSEDQUORUM that delegates tactical decisions to multiple commercial large language models.
- CAIRN works from file metadata—extracting embedded prompts, API endpoints, orchestration logic, and evasion text—and sorts results into three confidence tiers so researchers can hunt safely without running samples.
- CLOSEDQUORUM polls DeepSeek, Qwen, Mistral, and Google Gemini to vote on its next action and uses a fixed tiebreak order (DeepSeek, then Qwen, then Mistral, then Gemini) to choose steps like stealing credentials or crypto wallets.
- Talos’ static analysis showed the decision loop functions as designed but the analyzed sample contained placeholder API keys and a dummy webhook, so researchers did not observe it operating live or attribute its authorship.
- Researchers found about 20 other AI-integrated samples and say the discovery both raises questions about abuse of commercial LLMs and gives defenders a limited window to instrument metadata-based detections before autonomous C2 tactics spread.