Cisco has released Project Quarry as an open-source behavior detection engine. The tool is designed to monitor AI agents throughout their entire sessions to identify potentially malicious or unauthorized activity.
Unlike traditional security tools that focus on individual prompts, Quarry analyzes sequences of actions. It specifically flags suspicious patterns, such as an agent reading a sensitive file and subsequently sending data to an external endpoint.
The project aims to establish a shared language for agent behavior, similar to the roles Snort and YARA play in network traffic and malware detection. Rules in Quarry consist of a header, content, and metadata.
The GitHub repository currently provides a starter pack of rules and plugins for agents compatible with the Model Context Protocol, including Claude Code and Codex. Cisco emphasizes that these automated rules are intended to support, not replace, human oversight.
