Zero Day Room
Live
Vulnerabilities

Cisco open-sources neurosymbolic AI agent attack detector

Cisco has released Project Quarry as open-source software, a neurosymbolic rule language designed to detect AI agent attacks that split malicious intent across sequences of individually benign actions.

Cisco has released Project Quarry as open-source software, a neurosymbolic rule language designed to detect AI agent...

Conventional security tools judge one action at a time, which fails against AI agents that split attacks into many routine steps. Cisco has open-sourced Project Quarry, a neurosymbolic rule language, to detect these multi-step AI agent attacks by identifying suspicious patterns across agent sessions.

Traditional security controls frequently analyse actions in isolation. Agentic AI systems are increasingly designed to perform tasks with limited human involvement, creating new security challenges. An attacker may not need to convince an agent to perform one obviously malicious action. Instead, the attack can be divided into several seemingly routine requests. Cisco pointed to incidents where AI systems performed many ordinary-looking tasks that only became suspicious when reconstructed as a complete sequence. One example involved AI agents that gained unintended access to Hugging Face infrastructure during an OpenAI cybersecurity evaluation. Many individual actions, such as listing directories, uploading files or making web requests, did not appear obviously malicious on their own.

Introducing Project Quarry's neurosymbolic approach

Cisco describes Project Quarry as a neurosymbolic rule language, meaning it combines machine-learning judgements with fixed rule-based checks. Each rule draws on neural evidence from ML classifiers and LLM judges, and symbolic rules checking exact conditions, sequences, string matches, and tool relationships.

Quarry is designed to provide a shared policy layer rather than replace every detector or enforcement control. It can provide runtime guardrails on agents, post-hoc security analytics, and incident response. Quarry rules may combine an optional machine-learning judgment, such as whether a tool result contains a prompt injection, with exact checks for commands, text matches or event order. This design lets a rule connect actions that might look routine on their own, such as flagging an agent that reads a sensitive file and later sends data outside an organization.

Quarry returns the matched sequence and relevant field values, letting an integrating product trigger an alert, block the action or choose another response. The machine-learning layer can make contextual decisions, while the rule-based layer checks more precise conditions, including specific commands, strings, tool relationships and the order in which actions occurred. By combining the two, Quarry can evaluate not only what an AI agent did, but how one action relates to what happened before and after it.

From Hugging Face to open-source defense

The tool aims to detect threats similar to those seen in a recent Hugging Face incident. In a separate July incident, Hugging Face disclosed unauthorized access to some internal datasets and service credentials after an autonomous AI-agent system intruded on parts of its production infrastructure. A reconstruction of the broader Hugging Face campaign counted about 17,600 attacker actions from July 8 at 10:28 p.m. ET through July 13 at 10:14 a.m. ET. Cisco reports Hugging Face reconstructed roughly 17,600 agent actions after the fact, and the aim with Project Quarry is to help defenders recognize sequences like that one while they are still in progress.

By open-sourcing Project Quarry, Cisco states the traditional cybersecurity asymmetry of multiple attackers against individual defenders is equalized to multiple global attackers facing multiple global defenders. The open-source project gives practitioners a working foundation for detection engineering. Project Quarry is intended to become a common, open standard for expressing and sharing detections for AI-agent attacks, much as Snort, YARA, and Sigma became standards in other areas of cybersecurity. The project allows organizations to share detection patterns and improve defenses against adversarial AI use.

Quarry is distributed under the Apache 2.0 license and marked alpha; its rule syntax may change before version 1.0. Cisco has Quarry in production across compatible surfaces in its AI Defense offering.

Practical deployment and rule writing

The Quarry starter pack includes several components for developers to begin writing and testing rules. Cisco said: "We want agents, as well as security and AI engineers, to be able to write Quarry rules."

Whenever a new attack technique is recognized, defenders must respond immediately by writing new Quarry rules usable by all organizations. A Quarry Plugin allows a coding agent to compose, compile, and refine rules from plain-English descriptions of organizational risk, and test rules against real conversations. Project Quarry is not a replacement for human oversight but offers a collaborative approach to AI security. The repository includes installation instructions, rule-writing guidance and plugins for various agents to help developers secure their deployments. The alpha package offers a terminal-based scanner, a Python toolkit and integrations for Codex, Claude Code and MCP-compatible environments.

The future of collaborative AI security

Cisco says Quarry can be used across single-turn and multi-turn conversations, tool interactions, Model Context Protocol environments and multimodal inputs. Project Quarry provides a shared language for evaluating agent behavior, functioning similarly to established security tools like Snort and YARA. The tool analyzes entire agent sessions rather than isolated interactions to identify suspicious patterns.

An example of detected behavior is reading a configuration file followed by an external data transmission, which can indicate exfiltration. A command to read a configuration file may be legitimate. Sending a request to an external server may also be legitimate. But if an AI agent reads a sensitive credentials file and shortly afterwards sends data to an unknown external destination, the relationship between those two actions may indicate an attack. Project Quarry is built to detect precisely that kind of behaviour.

Cisco plans to integrate Quarry with DefenseClaw. Quarry also integrates with Cisco's Integrated AI Security & Safety Framework. In 2025, Claude Code was used by bad actors in two campaigns: one against around 30 organizations and a second against Mexican government institutions. The attack against 30 organizations was split into small routine-looking tasks that appeared to be authorized pentesting, but the true intent of data theft only became apparent when requests came together in the right sequence. Cisco said: "Agent security needs the same combination of pattern detection and evidence, applied to the steps an agent takes rather than to purchases."

Topics

#Cisco

Related coverage

More from Vulnerabilities