Cisco Open Sources Project Quarry to Detect Risky AI-Agent Behavior

Cisco released Project Quarry, an open-source behavior detection engine that helps security teams identify risky or unauthorized AI-agent activity across entire sessions. Rather than assessing isolated interactions, Quarry correlates sequences of actions, such as reading a sensitive configuration file and then transmitting data externally. It supports multi-turn large language model conversations, agent tool use, Model Context Protocol interactions, and multimodal inputs, with rules designed to flag behaviors including unauthorized tool calls and credential misuse.

The project provides plugins, starter rules, installation instructions, and rule-writing guidance, allowing organizations to share detection patterns through a common framework for evaluating agent behavior. Cisco suggested Quarry could flag activity resembling the OpenAI–Hugging Face incident, but researchers had not tested it against that incident’s data. Security teams can evaluate it as an additional monitoring layer, not a replacement for human oversight or chain-of-thought monitoring; its effectiveness depends on the detection rules deployed.