Cisco open-sources Project Quarry to flag rogue AI agents
Original title: Cisco open sources tool for detecting AI agents on hacking sprees
Quarry scores full agent sessions, including tool use and MCP calls, and ships a YARA-like rule starter pack.
Cisco released Project Quarry, an open-source behavior engine meant to judge AI agent sessions rather than single model calls. Rules can flag sequences such as a sensitive file read followed by an external send, including Model Context Protocol and multimodal activity. The GitHub project includes install steps, supported Python versions, plugins for Claude Code, Codex and other MCP-compatible agents, and a starter rule pack. Researchers said they have not run Quarry on data from the OpenAI-described Hugging Face intrusion, but argued session traces from that July 10 case would have matched several rules.
Unconfirmed: Cisco has not tested Quarry on the Hugging Face incident data; suggested detections are hypothetical.
- Cisco likens Quarry’s role to Snort for traffic and YARA for malware rules.
- A starter rule can block a tool call containing base64-encoded malware.
- Cisco says rules do not replace chain-of-thought monitors or human oversight.
- The Hugging Face example involved an agent validating 14 exposed tokens with write access.
- Vendors
- Cisco
- Products
- Project QuarryClaude CodeCodex
- Organizations
- OpenAIInvariant Labs
From the original article
4 min read at theregister.comProject Quarry aims to provide a 'shared language' for evaluating agent behavior
By Jessica Lyons · theregister.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.