What's Happening?
Cisco has open-sourced Project Quarry, a neurosymbolic rule language designed to detect AI agent attacks that are hidden within a series of seemingly harmless steps. Traditional security controls often fail to identify these sophisticated attacks because
they evaluate single actions in isolation, missing the malicious intent that only emerges when a sequence of benign actions is viewed collectively. Quarry combines machine learning judgment with rule-based checks, covering single-turn and multi-turn large language model (LLM) conversations, agents, tool use, Model Context Protocol (MCP) interactions, and multimodal inputs. Each rule in Quarry draws on neural evidence from ML classifiers and LLM judges, alongside symbolic rules that check exact conditions, sequences, string matches, and tool relationships. It functions as a shared policy layer, integrating with Cisco’s existing Integrated AI Security & Safety Framework to expand its capabilities from describing AI risk to actively detecting it in agent activity. The project aims to provide runtime guardrails on agents, facilitate post-hoc rule set security analytics, and enhance incident response.
Why It's Important?
The open-sourcing of Project Quarry is a significant development in cybersecurity, addressing a critical vulnerability posed by advanced AI agent attacks. As AI systems become more sophisticated, attackers can break down malicious activities into numerous small, routine-looking tasks that bypass traditional defenses. This was evident in 2025, when Claude Code was used in two campaigns against organizations and Mexican government institutions, where the true intent (data theft) only became apparent when individual requests were combined in the correct sequence. Quarry's ability to identify suspicious patterns that span individual permitted actions is crucial for protecting against these evolving threats. By open-sourcing the project, Cisco aims to equalize the traditional cybersecurity asymmetry, allowing a global community of defenders to collaborate and develop new rules in response to emerging attack techniques. This collaborative approach, combined with the ability for AI agents to write Quarry rules from plain-English descriptions, significantly accelerates the defense mechanism against AI-assisted attacks.
What's Next?
Project Quarry is intended to become a common, open standard for expressing and sharing detections for AI-agent attacks, similar to how Snort, YARA, and Sigma became standards in other cybersecurity domains. The open-source starter pack includes a readable Domain Specific Language (DSL) for conversations, tool use, and event relationships; an evaluation runtime; a command-line scanner; a Python SDK for application integration; and a Quarry plugin for Claude Code, Codex, and other MCP-compatible agents. The ongoing development will involve the continuous creation and refinement of new rules by both human security engineers and AI agents, responding to the dynamic landscape of AI threats. The goal is to enable defenders to recognize malicious sequences while they are still in progress and to preserve evidence for effective incident response. The success of Quarry will depend on widespread adoption and active community contribution to build a robust, collective defense against AI agent attacks.
Beyond the Headlines
The emergence of tools like Project Quarry highlights a fundamental shift in cybersecurity, moving from reactive defense against known threats to proactive detection of complex, multi-stage AI-driven attacks. This development underscores the increasing sophistication of cyber warfare, where AI is both a tool for attackers and a critical component of defense. The ethical implications of AI agents writing security rules are profound, raising questions about autonomous decision-making in cybersecurity and the potential for unintended consequences. Furthermore, the open-source nature of Quarry fosters a collaborative security ecosystem, challenging the traditional proprietary model of cybersecurity solutions. This collective intelligence approach could democratize advanced threat detection, making sophisticated defenses accessible to a wider range of organizations. Ultimately, Quarry represents a step towards a future where AI-powered defenses are essential for safeguarding digital infrastructures against equally intelligent adversaries, pushing the boundaries of what is possible in automated threat detection and response.
