Cisco open-sources Quarry to flag multi-step AI agent attacks
First merged summary.
Cisco has open-sourced Project Quarry, a rule engine that scores AI agent sessions, tool use, and MCP activity.
Cisco released Project Quarry, an open-source engine meant to judge full AI agent sessions rather than single model calls. Rules target intent that appears only across individually permitted steps, covering chats, tool use, Model Context Protocol activity, and multimodal inputs. It pairs machine-learning classifiers and LLM judges with symbolic checks of sequences, strings, and tool relationships, and ships a starter rule pack plus plugins for Claude Code, Codex, and other MCP-compatible agents. Cisco positions Quarry as a shared policy and evaluation layer for its Integrated AI Security and Safety Framework and says the rules do not replace chain-of-thought monitors or human oversight. Researchers said they have not run Quarry on the OpenAI-described Hugging Face intrusion but argued July 10 session traces would have matched several rules.
- Rules can flag a sensitive file read followed by an external send.
- A starter rule can block a tool call containing base64-encoded malware.
- Cisco likens Quarry to Snort for traffic and YARA for malware.
- A plugin lets coding agents draft and test rules from plain English.
- Hugging Face example: an agent validated 14 exposed tokens with write access.
- Cisco cites Claude Code campaigns split into routine-looking tasks.
What each outlet adds
beyond the shared facts- The RegisterSession scoring, YARA-like starter pack, and untested Hugging Face case.
- SecurityWeekNeurosymbolic design, framework role, and plain-English rule plugin.
- Vendors
- Cisco
- Products
- Project QuarryClaude CodeCodex
- Organizations
- OpenAIInvariant Labs
Coverage
oldest first · one row per article · UTCCisco open-sources Project Quarry to flag rogue AI agents
Quarry scores full agent sessions, including tool use and MCP calls, and ships a YARA-like rule starter pack.
Cisco open-sources Quarry to detect multi-step AI agent attacks
The neurosymbolic rule language is meant to flag intent that appears only across benign agent steps.