Cisco Open Sources Project Quarry To Detect Dangerous AI Agent Behaviour Across Entire Sessions

Cisco has open sourced Project Quarry, a detection language and engine designed to identify harmful behaviour across AI agent sessions, as security teams confront systems capable of reading files, executing commands and interacting with external services with increasing autonomy.

The project allows defenders to describe suspicious combinations of actions and evaluate recorded agent activity against those patterns. Cisco’s ambition is to establish a common approach to agent detection, comparable to the role shared rule formats have played elsewhere in cybersecurity.

The release addresses a difficult monitoring problem: an agent can perform several individually legitimate operations that collectively cross a security boundary. Access to a file may be authorised. So may an external request. Determining whether those…