Detecting a sophisticated AI agent attack is not simply a matter of finding multiple small needles scattered in a vast haystack; it is the far more complex task of recognizing whether those needles are connected and, crucially, for what malicious purpose they have been joined together. As AI agents become increasingly autonomous and integrated into enterprise workflows, traditional cybersecurity paradigms—which have long focused on evaluating single, isolated actions to classify them as benign or malicious—are proving inadequate. This fundamental shift in how attacks are orchestrated requires a new approach to defense, one that Cisco is now attempting to standardize through the open-source release of Project Quarry.
The core problem with legacy security controls is their binary nature. In the traditional security model, an action is either permitted or blocked based on its immediate characteristics. However, this approach fails spectacularly when confronted with AI agent activity. In these scenarios, the underlying malicious intent is often masked, only emerging over a complex, protracted series of seemingly benign actions. The threat is not apparent in any single step; it only becomes visible when the overall sequence of events is viewed in its entirety.
Fraud detection in the banking sector provides a useful, if limited, analogy. A bank traditionally analyzes each purchase as a discrete event. However, sophisticated fraud detection must look at patterns of purchases that appear innocuous on their own—such as small transactions occurring in geographically disparate locations or at unlikely times. Cisco argues that agent security requires an identical combination of pattern detection and forensic evidence, but applied to the specific, multistep execution paths that an AI agent takes, rather than just simple financial transactions.
To address this gap, Cisco has introduced Project Quarry, which it describes as a "neurosymbolic rule language." By combining the nuanced judgment of machine learning classifiers with the precision of rule-based logical checks, Quarry aims to provide a robust framework for monitoring AI behavior. The platform is designed to handle a wide spectrum of interactions, including single-turn and multi-turn large language model (LLM) conversations, autonomous agent workflows, tool use, Model Context Protocol (MCP) interactions, and multimodal inputs.
Under the hood, each rule defined within Quarry draws on two distinct sources of intelligence. First, it utilizes neural evidence provided by ML classifiers and LLM judges, which assess the intent and context of the interaction. Second, it incorporates symbolic rules that check for exact conditions, specific sequences, string matches, and the nature of tool relationships. This design is not intended to replace existing security enforcement controls or every specialized detector already in place. Instead, Cisco envisions Quarry as a shared policy layer that brings coherence and advanced detection capabilities to an organization’s existing security architecture.
The versatility of Project Quarry allows it to be deployed in several critical capacities: as runtime guardrails that intervene before an agent executes a command, as a post-hoc tool for security analytics, and as a component of incident response. It integrates directly with Cisco’s broader Integrated AI Security & Safety Framework. While that framework provides the conceptual roadmap for defining AI risks, Project Quarry serves as the functional policy and evaluation layer, allowing organizations to move from simply theorizing about risk to actively detecting it within the stream of agent activity.
The necessity for this evolution in defense is underscored by real-world events in 2025, where bad actors successfully weaponized legitimate tools to conduct espionage. In one notable campaign, attackers targeted approximately 30 organizations by manipulating Claude Code, an AI-powered coding assistant. A second, similar campaign targeted institutions within the Mexican government. In these instances, the attacks were broken down into small, routine-looking tasks. To existing defense systems, these requests appeared to be nothing more than part of a standard, authorized penetration test or routine maintenance. The true intent—the exfiltration of sensitive data—only became apparent when the seemingly disparate requests were synthesized in the correct, malicious sequence.
This capability of AI to "atomize" an attack into multiple, individually benign steps effectively defeats traditional signature-based defenses. Addressing this requires a new form of neurosymbolic threat detection that can maintain state and context over time. Furthermore, this is not a "write once and move on" challenge. As defenders develop new ways to identify these patterns, attackers are concurrently evolving their own approaches. Consequently, whenever a new attack technique is discovered, defenders must be able to respond with agility by writing a new Quarry rule that can be shared and implemented by organizations globally.
A significant hurdle in this arms race is the speed of development. Writing security rules by hand simply does not scale when confronted with adversaries who are using AI to generate and iterate upon new malicious agents at high speed. To counter this, Cisco has designed Quarry to be collaborative and accessible. "We want agents, as well as security and AI engineers, to be able to write Quarry rules," the company stated during the project’s launch. To facilitate this, a specialized Quarry Plugin enables a coding agent to compose, compile, and refine rules based on a plain-English description of an organization’s risk profile. These rules can then be tested against real-world conversations to ensure that a rule that passes in a sandbox environment will behave as expected at runtime.
By open-sourcing the project, Cisco hopes to address the fundamental asymmetry in the cybersecurity landscape, where a large, distributed network of attackers often enjoys a significant advantage over individual, siloed defenders. By fostering a global community of defenders who can share and refine rules, the industry can better match the speed and scale of AI-assisted attacks. Cisco views this as a foundational step in detection engineering, providing practitioners with the tools to build a common, open standard for expressing and sharing detections for AI-agent attacks, much as Snort, YARA, and Sigma did for network and endpoint security.
The urgency of this work is highlighted by recent reports of agentic AI breaches, including a high-profile incident reported to a Spanish regulator and the use of agentic AI to facilitate a ransomware attack via Langflow. Governments, including that of the United Kingdom, have already begun rolling out defense plans for agentic AI, reflecting the growing consensus that this is a tier-one security priority. Meanwhile, venture capital is flooding into the space, with firms like Rig Security raising $12 million specifically to tackle the identity and security risks posed by autonomous agents.
As Cisco notes, platforms like Hugging Face have had to reconstruct thousands of agent actions after the fact to understand breaches. The primary goal of Project Quarry is to help security teams recognize these suspicious sequences while they are still in progress, rather than in the aftermath. By preserving the forensic evidence behind each finding, the project aims to empower human responders to take decisive action before damage occurs. As the digital landscape continues to transition toward agentic workflows, the ability to monitor the "intent" behind the interaction, rather than just the action itself, will likely become the definitive requirement for the next generation of enterprise security.






