Cisco open sources Project Quarry to detect rogue AI agents

Cisco has released Project Quarry, an open-source behavior detection engine that evaluates AI agent sessions as a whole rather than isolated tool calls. Its Snort- and YARA-like rules can flag sequences such as reading a credentials file followed by an external send.

• Quarry covers single- and multi-turn LLM sessions, agents, MCP and multimodal inputs
• Rules capture action relationships, e.g. sensitive file read followed by external send
• Repo includes plugins for Claude Code, Codex and other MCP-compatible agents
• Cisco says rules complement, not replace, chain-of-thought monitoring and human oversight