Documentation overview
ATR (Agent Threat Rules) is an open standard for detecting threats to AI agents with rules. It has a specification and a versioned set of rules. It also has a test suite any engine can run to show it conforms.
- Specification
- 1.0.1 · Working Draft
- Rules
- 836 effective · 844 files on main · 8 inert
- Release
- npm 4.3.1
- Identifiers
- ATR-YYYY-NNNNN, never reassigned once published
- License
- MIT
- Governance
- Maintainer transition. The TSC is being seated.
How the documentation is organized
The normative text, how maturity and lanes work, and every change since release.
The ten threat categories. Each one lists its rules, their lanes and the crosswalks that come up most.
How rules map to OWASP, MITRE ATLAS, NIST AI RMF and regulation. ATR wrote these. They are not endorsements.
Engines, integrations, the conformance suite, and which part of ATR each adopter uses.
What a deployment can show an auditor, and how the auditor checks it against public material.
The ATD technique catalog and the public threat feed.
Who maintains ATR, how decisions are made, how to contribute and how to cite.
Common questions, who to contact about what, the privacy notice and the terms.
For machines
- llms.txtSite index for AI agents
- rules/index.jsonEvery rule with maturity, lanes and mappings
- rules/index.csvThe same data as CSV