Cisco open-sources CAIRN after finding malware that lets LLMs vote
CLOSEDQUORUM polls four LLM APIs to run itself with no human operator — a first. Plus: Meta hot-fixes a Muse zero-day that could backdoor a Mac.

Copy markdown
The toolkit: metadata forensics, no sandbox needed
CAIRN is MIT-licensed on GitHub (Cisco-Talos/CAIRN) and hunts AI-integrated malware by pulling “cognitive artifacts” — prompt templates, model endpoints, API-key prefixes — straight from file metadata, so you triage without ever running the binary. It ships 27 acquisition filters and 26 tiered YARA rules, and needs Python 3.11+ plus a VirusTotal Intelligence key.
CLOSEDQUORUM: a four-model committee runs the implant
The Windows implant Talos found has no human command server: it hands system context to a panel of up to four commercial LLMs — DeepSeek, Qwen, Mistral, Gemini — and lets them vote on its next move, with DeepSeek breaking ties. The winning call among credential theft, process injection or persistence runs on its own, and AES-256-GCM-encrypted loot is exfiltrated over Discord webhooks.
What it means if you ship on LLM APIs
The command channel is just ordinary calls to model APIs, so it blends into normal egress — the takeaway is to treat outbound requests to LLM endpoints as an attack surface. CAIRN's filter list doubles as a checklist: watch for stray API keys and langchain / litellm / ollama strings in your build artifacts, and log what your own agents actually send and receive.
Elsewhere: Meta hot-fixes a Muse zero-day that backdoored Macs
Patrick Wardle's Sept 21 proof-of-concept showed a hidden Muse setting (endo_voyager_dictation_endpoint) that let any local code reroute dictation to an attacker's server, lift the session token, and even drive Muse on a linked iPhone. Meta has since pushed a hot-fix — update the Mac app, or quit it if you're unsure you're patched.