Science & Tech · AI · 3 hrs ago
Cisco releases open-source tool to monitor AI agent behavior
Cisco has released Project Quarry, an open-source tool for monitoring AI agents.
The tool is designed to help security teams detect when AI agents take unauthorized actions.
Instead of judging each action on its own, Quarry examines the full sequence of actions in an agent session.
For example, reading a sensitive file may be legitimate, but sending its contents to an outside server afterward could signal data theft.
Quarry supports AI conversations, tool use and interactions through the Model Context Protocol, and lets teams write rules to flag suspicious patterns.
Cisco says the project is not a replacement for human oversight, but can help people investigate potential threats.
Its code repository includes setup instructions, plugins and guidance for writing detection rules.
Cisco has open-sourced Project Quarry, a tool designed to monitor AI agent behavior and help detect unauthorized actions.
The tool examines entire agent sessions to identify suspicious sequences that may not be apparent from individual actions.
It supports conversations, tool use, Model Context Protocol interactions and multimodal inputs.
Cisco says the project is intended to help security teams share detection rules and identify threats, not replace human oversight.
Its repository includes installation instructions, plugins, configuration examples and guidance for writing rules.
- Who
- Cisco released Project Quarry, an open-source behavior detection engine.
- What
- The tool monitors AI agent sessions to help identify suspicious behavior and unauthorized actions.
- When
- October 28, 2026.
- Where
- Not stated.
- Why
- To help security teams detect threats by examining the context and sequence of an agent’s actions.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- Project name
- Project Quarry
- Release
- Open-sourced by Cisco
- Session analysis
- Evaluates entire agent sessions
- Supported interactions
- Tool use and Model Context Protocol interactions
- Repository contents
- Plugins, installation instructions and rule-writing guidance
3 sources
What people are saying
But what is missing is the entire context under which something may occur.
One tool call may seem in-scope until you realize that that next step is taking that information and sending it to an attacker-controlled inbox or server.