Cisco's Quarry Targets Multi-Step AI Agent Attacks That Evade Traditional Tools
Written by AI from the linked story. Not reviewed by an editor. How this works
AI-powered agents can execute attacks through sequences of individually harmless actions, each of which passes standard security checks while collectively achieving a malicious goal. Traditional security tools fail here because they evaluate events in isolation rather than tracking behavioral patterns over time. This represents a fundamental blind spot in conventional logging and monitoring stacks, which were not designed for the emergent, multi-step threat model that AI agents enable. Cisco's open-source Quarry project addresses this gap by combining machine learning with rule-based logic to detect suspicious sequences — highlighting how defenders must evolve their monitoring capabilities to match attacker sophistication.
Tactical Insight
Immediate actions
- Deploy or evaluate neurosymbolic or sequence-aware detection tools (such as Cisco's open-source Quarry) to monitor AI agent workflows holistically.
- Audit existing SIEM and logging configurations to ensure they correlate multi-step events, not just individual alerts.
Long-term improvements
- Establish a behavioral baseline for all AI agents and automated systems operating in your environment so anomalous action sequences are detectable.
- Integrate AI-specific threat models into your vulnerability management program, explicitly covering agentic attack surfaces.
- Enforce least-privilege access for AI agents so that even a compromised sequence has limited blast radius.
Detection measures
- Implement chain-of-action logging that records the full context and sequence of AI agent decisions, not just individual API calls.
- Schedule regular red-team exercises simulating multi-step AI agent attack chains to validate detection coverage.