NIST AI RMF defines the process.
ATR makes it detectable.
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
Each of GOVERN, MAP, MEASURE and MANAGE has ATR rules under it. The breakdown by subcategory is below.
AI RMF is a governance framework. It tells you which risks to GOVERN, MAP, MEASURE and MANAGE, but it detects nothing on its own. ATR writes compliance.nist_ai_rmf metadata into the rules themselves. Each rule declares which subcategory it serves, so every function in the framework has detection under it that actually fires. This first shipped on 2026-05-09 with v2.1.0, when all 330 rules then in the corpus carried the mapping. As the corpus grew to 836 rules, the mapping grew to cover every one.
The mapping is not an “aligned with NIST” claim. Each one cites the detection element the rule actually relies on (a regex, token or signature), and you can check it against the YAML line by line. Anyone can download and audit it, and any auditor can refute it.
The framework itself is prose. For a machine to read it, someone has to put it into a structured format. The ATR maintainers turned NIST AI RMF into an OSCAL-format community catalog, with 72 controls and 176 cross-reference links, published under CC0 at Agent-Threat-Rule/ai-rmf-oscal-catalog. It is a community contribution, NOT a NIST publication.
Separately, two example AI RMF profiles were offered to NIST's usnistgov/oscal-content repository as PR #338. In June NIST asked that it wait for NIST's own AI RMF catalog. NIST added that catalog to its develop branch in July, and the PR was reworked against it in August. It is still open, waiting for NIST review. Whether to use any of it is NIST's call.
The NIST AI Risk Management Framework (AI RMF 1.0 plus the GenAI Profile) is NIST's framework for managing AI risk, meant for voluntary use. It splits AI risk management into four functions: GOVERN, MAP, MEASURE and MANAGE. That is the language of governance. It tells you what to manage, not how to catch the risk inside one SKILL.md file or one MCP tool description.
ATR fills that layer with per-rule metadata that is MIT-licensed, open source and reproducible. The AI RMF goes from a governance framework to a detection layer you can verify. A government, a SOC or an auditor can download the YAML and check, rule by rule, which subcategory each one maps to, why, and which detection element backs it.
The mappings land on 25 subcategories, and none of the four functions (GOVERN, MAP, MEASURE, MANAGE) is empty. Every part of the governance process has rules under it that fire. One rule can map to several subcategories, with primary and secondary strength. The table lists the count for each subcategory.
MS.2.7 is cited most, 665 times. One rule can map to several subcategories.
The mapping lives in the rule, not on a slide. Every rule's NIST mapping states which subcategory it serves and why. Below is the actual metadata from ATR-2026-00118 (Approval Fatigue Exploitation):
compliance:
nist_ai_rmf:
- subcategory: GV.6.1
context: Approval fatigue exploitation manipulates
human-in-the-loop oversight by overwhelming operators
with rapid permission requests or minimizing
dangerous actions; GV.6.1 requires data and oversight
governance policies that preserve meaningful human
review rather than enabling bulk auto-approval of
risky tool calls.
strength: primaryThe context field says why this rule belongs under GV.6.1. Not generic “governance”, but the specific path by which an approval-fatigue attack breaks oversight policy. Every rule is written up this way.
The mapping pipeline has three stages: LLM-assisted batch generation, per-rule QA, and an atomic patch. All of it is open source and reproducible.
- InputThe 330 ATR rule YAMLs in v2.1.0 (detection patterns, test cases, existing metadata), the NIST AI RMF 1.0 reference, the GenAI Profile, and hand-written 5-shot examples. Rules added since v2.1.0 get their mappings from the same batch pipeline.
- Batch generator
scripts/expand-nist-mapping.ts: Claude Opus with a 5-shot prompt and structured output. Each rule gets at least 1 primary and 0 to 3 secondary subcategory mappings, each with its own context field. Every subcategory ID is checked strictly against the RMF reference, so none are made up. - Atomic patcher
scripts/apply-nist-mapping.ts: Reads each proposal YAML and patches the compliance.nist_ai_rmf block in the matching rule YAML, writing atomically (tmp, then rename). Every patched YAML still parses (0 failures out of 261). Mappings a person already curated are never overwritten. - Scale261 new mappings on top of the 69 that already existed, covering all 330 rules then in the v2.1.0 corpus.
- ProvenanceEvery rule's proposal YAML is kept under proposals/nist/. Anyone can re-run the pipeline, compare the output, and audit the reasoning behind a mapping.
NIST has several public lines of work on AI agent security: CAISI’s request for information on securing AI agent systems (docket NIST-2025-0035), the COSAiS SP 800-53 control overlays for single-agent and multi-agent systems, and the NCCoE Software and SI Agent Identity and Authorization project. ATR is an independent open-source project and has no role in any of it. NIST has not selected or endorsed ATR.
The mappings on this page are rationales drafted with LLM help (see Methodology). They are claims about relevance, not measurements. Detection performance is measured separately.
- RFI docket: NIST-2025-0035 (CAISI Issues Request for Information About Securing AI Agent Systems)
- Related project: NCCoE Software and SI Agent Identity and Authorization, on identity and authorization for agents. Detection is a separate layer, and it needs that identity context.
- Benchmark figures, reported per lane, with closed-book and in-sample caveats: README §8 Evaluation · DOI 10.5281/zenodo.19178002
The mapping is open metadata, not a closed spec you can only read. Every rule's RMF mapping is public YAML on GitHub. Fork it, challenge it, or open a PR to fix its strength or context.
Broader context: Sovereign AI Defense: Open Call
