Cisco open sources Project Quarry for AI agent behavior detection

Project Quarry builds on a lineage of detection tools Cisco already maintains. The company positions Quarry as providing a 'shared language' for agent behavior evaluation, similar to what Snort did for malicious network traffic and YARA for rule-based malware identification.

The design stems from a gap Cisco researchers see in current detection methods. Most tools check for specific interactions, such as a prompt injection attempt or a call to an attack tool, but miss the context under which actions occur. Quarry instead analyzes the entire agent session, allowing rules to correlate events. For example, a rule can produce a finding when a sensitive file read is followed within a defined window by an external send, preserving evidence from both actions.

Cisco's team acknowledges that rules alone are not a substitute for chain-of-thought monitors or other human oversight, but they help record patterns that other teams can review, test, and apply to their own agents. The open-source approach is intended to let defenders share the slices of the problem they see, as Chang put it, rather than wait for a complete solution.