Skip to content
Agent Threat Rules (ATR)中文
Compliance coverage

Every rule is tagged against 6 frameworks.

Regulations ask you to manage AI agent risk. They do not say how to prove it on a real artifact. Each of ATR's 836 rules carries six framework mappings: OWASP LLM, OWASP Agentic, MITRE ATLAS, NIST AI RMF, EU AI Act and ISO 42001. CI checks that every mapping is valid and that every rule has one. Each rule's detection runs on a SKILL.md file, an MCP tool description or an agent config, and when it fires it provides detection evidence supporting the requirements it maps to. That does not mean every requirement in a framework has a detection, or that a rule satisfies the requirement. SAF-MCP is mapped per technique, as candidate mappings (55/78, not validated by SAF). All of the metadata is MIT-licensed. You can download it and audit it.

Some framework mappings are known to be wrong and are being redone; detection is not affected. Until the corrections ship, treat the mappings as candidate relevance, not as audit evidence. See the errata

Articles 9 to 15 of the EU AI Act apply to high-risk AI systems, as classified under Article 6 and Annex III. The obligations sit mainly with the provider. A deployer has its own duties under Article 26. An ATR mapping to one of these articles is detection evidence an operator may cite in its own risk management. It is not compliance. Some rules also cite CEN-CENELEC JTC 21 standards. Citing a standard here does not create a presumption of conformity under Article 40.

For procurement teams

There is no packaged download yet. Each rule's mappings live in its own YAML file, and the identifiers each framework allows live in data/compliance-frameworks/ in the repo. The coverage policy says what a mapping means and what it does not. Because ATR is an open standard, a reviewer doesn't have to trust it. They can check every mapping themselves.

Every mapping is ATR's own crosswalk. None is endorsed by the bodies named here.

OWASP LLM Top 10 (2025)
Every rule

All 836 rules carry an OWASP LLM Top 10 (2025) reference.

OWASP Agentic Top 10
Every rule

All 836 rules carry an OWASP Agentic Top 10 reference, and all 10 ASI risk categories are covered.

MITRE ATLAS
Every rule

All 836 rules carry a MITRE ATLAS technique reference, and the rule explorer groups them by tactic. The ATLAS crosswalk is ATR's own.

NIST AI RMF
Every rule

All 836 rules carry NIST AI RMF subcategory mappings across the GV, MP, MS and MG functions. A community-written OSCAL catalog (CC0) is self-published. Two example AI RMF profiles sit in an open pull request (#338) to NIST's OSCAL content repository, waiting for NIST review. This is not a NIST endorsement or adoption.

EU AI Act
Every rule

Each of the 836 rules cites at least one of Articles 9, 10, 12, 13, 14 and 15. Those articles cover high-risk AI systems only. The citation names the duty the rule's runtime detection can give evidence for. It does not mean a system that runs ATR complies with the Act.

ISO 42001
Every rule

All 836 rules map to AI management system clauses (6.2, 8.1–8.4, 9.1).

SAF-MCP
55/78 techniques

Counted per technique: 55 of 78 active SAF Framework Model v2 techniques have at least one reviewed ATR rule, as a partial or adjacent match. These are candidate mappings. SAF has not validated them, and SAF validates outside mappings in its own registry.

Check it rule by rule

'Aligned with a framework' is an assertion. The regex that detects the attack is evidence. ATR's compliance mappings live in each rule's YAML. They cite the exact regex or token that detects the attack, never a generic 'aligned with the framework'. Anyone can put a mapping next to the thing it claims to detect and check it.

Browse raw rule YAML on GitHub

Think a mapping is wrong? It's an open standard. Fork it, challenge it, and open a PR or an issue to fix the mapping.

Open an issue to challenge a mapping