Cisco open-sources Quarry to detect multi-step AI agent attacks
Original title: Cisco Open Sources Quarry to Catch AI Agent Attacks Hidden in Harmless Steps
The neurosymbolic rule language is meant to flag intent that appears only across benign agent steps.
Cisco open-sourced Project Quarry, a neurosymbolic rule language for detecting AI-agent attacks whose intent emerges across individually permitted actions. It combines machine-learning evidence with symbolic checks of sequences, strings, and tool relationships, covering chats, agents, tool use, MCP, and multimodal inputs. Cisco positions it as a shared policy and evaluation layer for its Integrated AI Security and Safety Framework, with runtime guardrails, analytics, and incident response uses. A plugin lets coding agents draft and test rules from plain English.
- Quarry pairs ML classifiers and LLM judges with exact symbolic checks.
- It covers single-turn and multi-turn LLM conversations and MCP interactions.
- A plugin targets Claude Code, Codex, and other MCP-compatible agents.
- Cisco cites Claude Code campaigns split into routine-looking tasks.
- Vendors
- Cisco
- Products
- Project QuarryClaude Code
From the original article
4 min read at securityweek.com
Cisco open sourced Project Quarry, a rule language for detecting AI agent attacks that only emerge across sequences of benign actions. The post Cisco Open Sources Quarry to Catch AI Agent Attacks Hidden in Harmless Steps appeared first on SecurityWeek.
By Kevin Townsend · securityweek.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.