Cisco releases Project Quarry to spot rogue behaviors across AI agent sessions

Cisco has open sourced Project Quarry, a behavior detection engine and rule language designed to track AI agent activities across entire sessions. Announced in late October, the project aims to detect attacks that surface through sequences of seemingly benign actions rather than isolated commands. The tool's GitHub repository offers installation instructions, supported Python versions, a starter pack of rules, and plugins for Claude Code, Codex, and other Model Context Protocol-compatible agents. [1][2]

Project Quarry inspects single- and multi-turn conversations, tool calls, Model Context Protocol interactions, and multimodal inputs using rules patterned after Snort and YARA syntax. Cisco researchers explain that evaluating full session context is critical, as a routine task like reading a configuration file can turn malicious if immediately followed by sending that data to an external server. Rules can identify anomalies such as base-64 encoded malware, unapproved internet access from sandboxes, or the extraction and misuse of sensitive credentials. [2]

According to Cisco researchers, while rules alone cannot replace human oversight or chain-of-thought monitoring, providing an open standard allows security teams to share and apply detection logic collectively. Cisco pointed to incidents such as the July 10 Hugging Face breach, suggesting that evaluating complete multi-step sessions can expose rogue behaviors like testing and sharing exposed write tokens before major damage occurs. [2]