NewsPublished 29 Oct 2026

Cisco Open Sources Quarry to Catch AI Agent Attacks Hidden in Harmless Steps

Source SecurityWeek

cisco

Description

Cisco has open sourced Project Quarry, a neurosymbolic rule language designed to detect AI agent attacks that manifest only across sequences of individually benign actions. Traditional security controls often fail to detect such attacks because they evaluate actions in isolation rather than as connected sequences. Quarry combines machine learning classifiers and rule-based checks to identify suspicious patterns in AI agent behavior over time. It supports runtime guardrails, post-hoc analytics, and incident response, integrating with Cisco's AI Security & Safety Framework. The project aims to become a shared standard for detecting AI-agent threats and to enable defenders worldwide to rapidly develop and share detection rules, countering attackers' use of AI to create sophisticated multi-step attacks.

Impact

The impact of this development is primarily defensive: it enhances the ability of organizations to detect sophisticated AI-driven attacks that evade traditional security by distributing malicious intent across multiple benign actions. This capability is critical as attackers increasingly use AI to orchestrate complex, multi-step attacks that are difficult to detect with conventional methods. While no direct exploit or vulnerability is described, Quarry addresses a significant gap in AI agent threat detection, potentially reducing the risk of data theft and other harms from such attacks.

Defensive guidance

Project Quarry is an open-source detection tool designed to augment existing security controls by identifying suspicious AI agent activity across sequences of actions. Organizations should consider integrating Quarry into their AI security frameworks to improve detection capabilities. Since Quarry is open source and actively maintained, defenders can write and share new detection rules as new attack techniques emerge. There is no indication that immediate patching or other remediation is required; rather, the recommendation is to adopt and customize Quarry rules to address evolving AI agent threats.

Summary

Project Quarry is a neurosymbolic rule language open sourced by Cisco to detect AI agent attacks that only become apparent when analyzing sequences of benign actions collectively. It combines neural evidence from ML classifiers and LLM judges with symbolic rules to evaluate single-turn and multi-turn conversations, tool use, and multimodal inputs. Quarry addresses the challenge where AI attackers split malicious intent into multiple harmless steps, defeating traditional security controls that assess actions individually. The project includes a DSL for defining rules, an evaluation runtime, integration SDKs, and a plugin for AI coding agents to compose and test rules. By open sourcing Quarry, Cisco aims to equalize the cybersecurity asymmetry by enabling global defenders to rapidly create and share detection rules, keeping pace with evolving AI-enabled attack techniques.

References

securityweek.com/cisco-open-sources-quarry-to-catch-ai-agent-attacks-hidden-in-harmless-steps/ (opens in a new tab)

Community reviews

Loading reviews…

Description

AI analysisMachine-generated, updated 29 Oct 2026, 14:32 UTC

Technical details

References

Community reviews

Related threats