Cisco Open-Sources Quarry for Detecting Suspicious Multi-Step AI Agent Behaviors

Networking and cybersecurity vendor Cisco has open-sourced an AI agent behavior detection tool named Project Quarry, releasing the source code and documentation on GitHub. Quarry is designed to analyze tool calls, execution outcomes, and the sequence of operations performed by AI agents during multi-step tasks, assisting security teams in identifying suspicious activities. According to the documentation, Quarry allows security personnel to define detection conditions via a dedicated rule language and can incorporate AI models for semantic checks, such as identifying prompt injection within tool responses or detecting deviations from assigned tasks. Cisco stated that Quarry is already utilized within its AI Defense product. The documentation also notes that organizations adopting Quarry must maintain detailed logs of agent operations, as keeping only user prompts and final outputs is insufficient for reconstructing the execution process. Organizations can test rules against historical logs to reduce false positives or integrate the tool into runtime security verification pipelines.

Key takeaways

• Cisco open-sourced Project Quarry on GitHub, offering a dedicated rule language and engine to monitor AI agent actions.
• The tool analyzes tool calls, results, and operation sequences, supporting AI-assisted checks for prompt injections and task drift.
• Cisco reported using Quarry in its AI Defense product; deployment requires retaining granular logs of agent interactions.