Skip to content

ATR threat detection: 108 rules as probes + detector for AI agent security #1677

Description

@eeee2345

ATR (Agent Threat Rules) has 108 open-source detection rules for AI agent threats. Cisco AI Defense ships 34 of them in production. Proposing a dual integration with Garak:

As detector: garak/detectors/atr.py — scan LLM outputs for credential leaks, prompt injection artifacts, tool poisoning indicators. 108 regex patterns, <1ms per check, zero API keys.

As probe source: ATR's 824 verified test cases (true_positive payloads) are ready-made attack inputs covering prompt injection, tool poisoning, credential exfiltration, and 6 more categories. Each payload is labeled with expected behavior.

This gives Garak users red+blue in one tool — attack with ATR payloads, detect with ATR rules.

Numbers: 96.9% recall on SKILL.md benchmark (498 samples), 99.7% precision on PINT (850 samples), 0% FP on 53K real-world skills.

Would the team be open to a PR? Happy to follow existing detector/probe patterns.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions