Cisco Open Sources Quarry to Detect Hidden AI Agent Attacks

Cisco has open sourced Project Quarry, a neurosymbolic rule language designed to detect AI agent attacks that only reveal themselves when a sequence of benign actions is viewed together. The move aims to counter attacks that leverage AI agents by recognising patterns that span multiple steps rather than judging single actions in isolation.

Quarry is described as a shared policy layer that combines machine learning judgments with symbolic rules to assess conversations, tool use, Model Context Protocol interactions and multimodal inputs.

Cisco positions Quarry as a means to provide runtime guardrails on agents, post-hoc rule security analytics, and incident response, integrating with its Integrated AI Security & Safety Framework.

The Quarry starter pack includes a readable DSL for conversations and tool relationships, an evaluation runtime with configurable safety limits, a command-line scanner, a Python SDK, optional model operation configurations for semantic signals, a Quarry plugin for Claude Code, Codex and MCP-compatible agents, plus rule metadata, examples and a workflow for testing rules against positive and benign datasets.

The company notes Quarry is intended to become a common open standard for expressing and sharing detections of AI‑agent attacks, likening its ambition to Snort, YARA and Sigma in other security domains. Cisco cites past incidents where agent-like activity split into seemingly routine tasks only revealed malicious intent when sequences coalesced, emphasising the need for rapid rule development and a scalable, community-driven approach.