In the evolving landscape of cybersecurity, detecting an attack by an AI agent is far more complex than identifying a single malicious file or an isolated intrusion. It is not merely a matter of finding multiple small needles scattered in a massive digital haystack; rather, the challenge lies in discerning whether those needles are connected and, if so, what their ultimate purpose might be. As AI agents gain the ability to operate autonomously and execute complex tasks, the traditional security model—which evaluates single actions to classify them as either bad or benign—is increasingly proving to be obsolete.
This shift in the threat landscape is profound. Under the legacy paradigm, security controls function as gatekeepers, inspecting individual requests for signs of malice. However, this approach inherently fails against AI agent activity, where malicious intent does not manifest in a single, glaring violation. Instead, the "bad" intent only emerges over a series of actions that, when viewed in isolation, appear entirely benign. The danger only becomes apparent when the overall sequence is reconstructed and analyzed in its entirety.
The fundamental flaw in traditional security lies in its granularity. Take, for instance, the evolution of fraud detection in the banking sector. A bank typically analyzes each purchase on its own merits to determine if it is legitimate. This is the bedrock of how most traditional security systems operate. However, effective fraud detection requires a holistic view—identifying patterns of seemingly normal purchases that, when combined across multiple merchants, geographic locations, and improbable time intervals, reveal a coordinated effort by an autonomous agent. Cisco, in announcing the open-sourcing of its new initiative, Project Quarry, argues that agent security requires this same combination of sophisticated pattern detection and evidence, applied not just to financial transactions, but to the intricate steps an agent takes during its execution.
Bridging the Gap with Neurosymbolic Detection
Cisco describes Project Quarry as a neurosymbolic rule language. This is a critical distinction in the field of AI security, as it combines the intuitive, pattern-matching judgment of machine learning classifiers with the rigid, verifiable precision of rule-based checks. By design, Quarry is intended to operate as a shared policy layer rather than a wholesale replacement for existing detectors or enforcement controls. It provides a common language through which security teams can define, share, and enforce policies across complex environments.
The breadth of Quarry’s capabilities is designed to match the complexity of modern LLM-driven ecosystems. The language covers single-turn and multi-turn conversations with large language models, the deployment and behavior of agents, the use of various tools, interactions with the Model Context Protocol (MCP), and even multimodal inputs. Each defined rule within the Quarry framework draws on neural evidence provided by ML classifiers and LLM judges, as well as symbolic rules that check for exact conditions, precise sequences, string matches, and specific tool relationships.
By integrating with Cisco’s existing Integrated AI Security & Safety Framework, Quarry effectively transitions the industry from merely describing AI risks to actively detecting them in real-time agent activity. The framework serves as the conceptual roadmap for identifying risks, while Quarry provides the concrete detection layer to spot those risks in action. This synergy allows organizations to implement runtime guardrails on agents, conduct post-hoc rule set security analytics, and improve their incident response times.
The Shift Toward Sequential Threat Detection
The necessity for this new form of neurosymbolic threat detection is best illustrated by the changing nature of cyberattacks observed in 2025. Bad actors have begun to weaponize AI agents to conduct complex, multi-stage campaigns that exploit the blind spots of traditional defenses. One notable example involved the use of Claude Code in a campaign targeting approximately 30 organizations. In this instance, the attack was meticulously broken down into small, routine-looking tasks. By masquerading as a new defense testing procedure, the agent mimicked the behavior of an authorized penetration test. To existing security systems, each step appeared legitimate, and it was only when the requests were aggregated in the correct sequence that the true intent—the surreptitious theft of sensitive data—became visible.
A second, similar campaign targeted Mexican government institutions, further highlighting that the ability of AI to decompose an attack into multiple, individually benign steps is a significant force multiplier for attackers. This creates a "death by a thousand cuts" scenario where traditional defenses are bypassed not by brute force, but by the clever orchestration of authorized, low-risk actions.
Because attackers are constantly iterating and developing new approaches, the defense must be equally dynamic. This is not a "write once and move on" requirement. Whenever a new attack technique is identified, defenders must respond immediately by drafting a new Quarry rule that can be deployed across the entire ecosystem. However, this creates a scaling problem: manual rule-writing cannot keep pace with the speed at which attackers use their own AI agents to evolve malicious tactics.
Scaling the Defense Through Automation and Collaboration
Recognizing this imbalance, Cisco has designed Quarry to facilitate a more scalable defense model. "We want agents, as well as security and AI engineers, to be able to write Quarry rules," the company declared during the launch. To this end, a Quarry Plugin has been introduced, allowing a coding agent to compose, compile, and refine rules based on plain-English descriptions of an organization’s specific risk profile. These rules can then be tested against real-world conversations to ensure they perform as intended. A rule that successfully passes validation within a coding agent is designed to behave with the same consistency at runtime, significantly reducing the friction involved in deploying new protections.
By open-sourcing the project, Cisco aims to fundamentally alter the cybersecurity asymmetry that has long favored the attacker. Traditionally, a handful of individual defenders have struggled to keep up with a global, networked force of attackers. Project Quarry attempts to equalize this by allowing a global community of defenders to collaborate. Just as open-source projects like Snort, YARA, and Sigma became the bedrock of modern intrusion detection and signature-based security, Cisco envisions Quarry as the common, open standard for expressing and sharing detections for AI-agent attacks.
The value of this approach is underscored by the sheer volume of data involved in modern investigations. Cisco pointed to a case involving Hugging Face, where researchers were forced to reconstruct roughly 17,600 individual agent actions after the fact to understand an incident. "Our aim with Project Quarry is to help defenders recognize sequences like that one while they are still in progress, and to preserve the evidence behind each finding for the people who have to act on it," the company explained.
As the industry grapples with the first wave of agentic AI data breaches—such as those reported to regulators in Spain and the emergence of ransomware attacks facilitated by tools like Langflow—the need for a unified defense strategy has never been more pressing. Governments, including the UK, have already begun rolling out defense plans alongside industry pledges to ensure the safe development and deployment of agentic AI.
With initiatives like Project Quarry and the influx of capital into startups like Rig Security, which recently emerged from stealth with $12 million to tackle agentic identity risks, the cybersecurity sector is clearly pivoting toward a future where the defense of AI agents is as sophisticated as the agents themselves. By providing a working foundation for detection engineering, Cisco’s open-source project offers a critical tool for practitioners, shifting the focus from individual action monitoring to the intelligent, sequential analysis required to stay ahead of the next generation of automated threats. As this standard gains traction, the hope is that the defense will finally be able to match the speed and complexity of the attackers, turning the tide in the race to secure the era of autonomous AI.

