A new rule language aims to catch agent attacks that only become visible across a whole sequence of individually ordinary actions.
SecurityWeek reports that Cisco has open sourced Project Quarry, a rule language built to catch AI agent attacks that only become visible across a sequence of individually harmless steps. Quarry is a neurosymbolic system, pairing machine learning classifiers and LLM judges with exact checks on conditions, sequences, strings and tool relationships, and it covers LLM conversations, agents, tool use, MCP interactions and multimodal inputs.
The release includes a readable rule language, an evaluation runtime, a command line scanner, a Python SDK and a plugin for Claude Code, Codex and other MCP-compatible agents that can draft and test rules from a plain English description of an organization's risk. Cisco positions Quarry as the policy and evaluation layer for its existing AI security framework and hopes it becomes an open standard for sharing agent-attack detections, similar to Snort, YARA and Sigma.
Why it matters: as enterprises hand more work to autonomous agents and MCP-connected tools, single-step guardrails miss attacks that only add up across a whole session. A shared, open rule language for agent behavior gives security and platform teams a common way to detect and share those patterns, the same governance enterprise AI deployments need as they scale.