What Cisco's Project Quarry plugin for Claude Code does (theainewsreport.com)

Cisco has announced the open-sourcing of Project Quarry, a neurosymbolic rule language designed to identify risky behaviors in AI agents. Released on October 28, 2026, Quarry enables developers to define rules that check the sequences of agent actions, effectively functioning like Snort or YARA for agent conversations. The core mechanism involves compiling rules to finite-state machines, allowing specific patterns—such as a read action followed by a network call—to trigger alerts while ignoring reversed sequences. A plugin for Claude Code and Codex integrates tools for rule writing and conversation scanning, but notably does not monitor live agent actions. The significance of Quarry lies in its potential to enhance security in AI deployments by providing a structured way to analyze conversation logs for harmful patterns. As AI agents increasingly interact with sensitive information, the ability to preemptively identify risky behaviors is crucial. The plugin facilitates local rule execution on exported transcripts, empowering developers to create and test security measures without impacting the live operation of agents. However, users are cautioned that Quarry should not be viewed as a runtime guard, as it requires manual intervention for transcript analysis, highlighting the need for robust security practices in AI development.