NEWS

Cisco Open Sources Quarry to Catch AI Agent Attacks Hidden in Harmless Steps

Cisco's Project Quarry is a neurosymbolic rule language for detecting AI agent attacks that only emerge across sequences of individually benign actions.

Dylan H.

News Desk

October 29, 2026
2 min read
Cisco Open Sources Quarry to Catch AI Agent Attacks Hidden in Harmless Steps

Overview

Cisco has open sourced Project Quarry, a rule language for detecting attacks on AI agents that only become visible across a sequence of individually harmless actions. Cisco calls it a neurosymbolic rule language, combining machine-learning judgment with deterministic rule checks.

Quarry covers single-turn and multi-turn LLM conversations, agents, tool use, Model Context Protocol (MCP) interactions, and multimodal inputs.


How Rules Work

Each rule can combine two kinds of evidence:

  • Neural evidence from ML classifiers and LLM judges
  • Symbolic checks such as exact conditions, sequences, string matches, and tool relationships

Cisco positions Quarry as a shared policy layer rather than a replacement for every detector or enforcement control. Intended uses are runtime guardrails for agents, post-hoc security analytics across rule sets, and incident response. It integrates with Cisco's Integrated AI Security & Safety Framework, which defines AI risks, while Quarry detects them in agent activity.


Starter Pack

ComponentPurpose
Readable DSLExpress conversations, tool use, and event relationships
Evaluation runtimeCompiles each rule on load, with configurable safety and resource limits
CLI scannerScan individual traces, directories, and datasets
Python SDKIntegrate into applications
Model operationsOptional configured models for semantic signals
Agent pluginFor Claude Code, Codex, and other MCP-compatible agents
Validation workflowTest rules against positive and benign datasets

The plugin lets a coding agent compose, compile, and refine rules from a plain-English description of an organization's risk. Cisco says a rule that passes in a coding agent should behave the same way at runtime.


Why It Matters

Many agent attacks, such as data exfiltration through a chain of permitted tool calls, never trip single-event detectors. Cisco wants Quarry to become a common, open standard for sharing AI-agent attack detections, in the way Snort, YARA, and Sigma did for their domains. Its success will depend on community adoption. The coverage reviewed did not state the license or repository location, so check Cisco's announcement before adopting it.


Sources