Cisco Open Sources Project Quarry for AI Agent Security

In a company blog post, Cisco introduced and open sourced Project Quarry, a rule language and detection engine for AI agent security. It evaluates complete agent sessions to identify harmful patterns that may not be apparent from individual actions.

Quarry combines machine learning judgments with exact rule based checks. It supports single turn and multiple turn model conversations, tool use, Model Context Protocol interactions, and image inputs. Rules can inspect events such as sensitive file access followed by data transmission to an external destination, then return the steps and fields that caused a match.

Security teams can use Quarry during model calls and tool runs, apply it to stored traces, or test newly identified attack patterns against earlier sessions. Systems receiving a finding can use it to trigger actions including alerts, blocking, redaction, quarantine, or logging.