What happened
On Oct 28 Cisco open-sourced Project Quarry, a neurosymbolic rule language plus an evaluation engine for detecting harmful or disallowed AI-agent behavior across multi-turn conversations, tool use, MCP interactions, and multimodal inputs. Rules combine ML/LLM-judge neural signals (e.g., prompt-injection in a tool result) with exact symbolic checks, and return evidence (steps, matched fields, masked secret excerpts) that products map to alert/block/redact/quarantine actions.
Why it matters
It aims to be a shared, testable policy layer for agent security analogous to Snort/YARA/Sigma, addressing attacks that only materialize across a sequence of individually permitted actions — a gap single-message classifiers and per-action authz controls miss.
Applicability
Security teams building agent runtime guardrails, post-hoc trace evaluation, and incident response should evaluate the OSS for session-level agent detection; vendors should consider adopting Quarry as the shared rule format.