Cisco open-sources Project Quarry, a detection language for AI agent behavior (blogs.cisco.com)

Cisco has announced the open-sourcing of Project Quarry, a detection language specifically designed to monitor AI agent behavior. As AI agents have begun to operate autonomously, traditional security measures fall short in tracking the multi-step actions they perform, which may seem benign individually but can pose significant security risks when assessed collectively. Quarry provides security teams with a structured way to define and detect harmful agent behavior by employing a neurosymbolic rule language that combines machine learning with exact rule-based checks. This enables teams to describe complex attack patterns, assess agent interactions, and extract relevant evidence of potential threats. The significance of Quarry lies in its ability to move beyond simple message classification to recognizing threat patterns that accumulate over a sequence of actions, reminiscent of fraud detection in banking. By allowing for the evaluation of both single and multi-turn interactions, Quarry serves as a crucial policy layer that enhances incident response and runtime security evaluations. It integrates seamlessly with Cisco's AI Security & Safety Framework, creating a shared language for describing AI risks and enforcing detection policies. This open-source tool is expected to facilitate collaborative efforts in the AI/ML community for more robust security measures against evolving threats posed by autonomous agents.