NIST AI RMF Generative Profile: Mapping Controls to Your LLM Stack

The Framework

NIST's AI Risk Management Framework (AI RMF 1.0, 2023) gave organizations a structured way to identify, measure, manage, and govern AI risks. The Generative AI Profile (NIST AI 600-1, July 2024 with 2025 updates) specialized the framework for generative AI risks.

By 2026, US federal contracts and many enterprise procurement RFPs reference RMF compliance. This piece maps the high-level controls to the parts of a real LLM stack.

The Four Functions

flowchart TB
 Govern[GOVERN<br/>policies, accountability, culture] --> Map
 Map[MAP<br/>context, risks, intended use] --> Measure
 Measure[MEASURE<br/>tests, metrics, evaluation] --> Manage
 Manage[MANAGE<br/>treat, monitor, incidents] --> Govern

The functions cycle. Each one is a section of the framework with measurable subcategories. The Generative Profile adds GenAI-specific risks and controls under each.

The 12 Generative-AI-Specific Risks

The Profile identifies risks that are heightened or unique to generative systems:

1. CBRN information / capabilities
2. Confabulation (hallucination)
3. Dangerous, violent, or hateful content
4. Data privacy
5. Environmental impacts
6. Harmful bias / homogenization
7. Human-AI configuration (over-reliance, automation bias)
8. Information integrity (misinfo, deepfakes)
9. Information security
10. Intellectual property
11. Obscene, degrading, abusive sexual content
12. Value chain / component integration

Most production LLM applications care most about confabulation, data privacy, harmful bias, IP, and security.

Mapping Controls to a Real Stack

flowchart LR
 Stack[LLM Stack] --> L1[Data Layer]
 Stack --> L2[Model Layer]
 Stack --> L3[Prompt + Tool Layer]
 Stack --> L4[Application Layer]
 L1 --> C1[Privacy controls,<br/>data classification]
 L2 --> C2[Provider attestations,<br/>model cards]
 L3 --> C3[Prompt guards,<br/>tool permissions]
 L4 --> C4[Human oversight,<br/>logging, eval]

Data Layer

• Classify training and inference data by sensitivity
• Document provenance and licensing
• Enforce data minimization in prompts (no PII unless required)
• Logging captures only what is necessary

Model Layer

• Use providers that publish model cards and safety evaluations
• For fine-tuned models, maintain your own model card
• Pin model versions; document version changes
• For self-hosted models, run the standard safety eval suite

Prompt + Tool Layer

• Input guards (prompt injection detection)
• Output guards (PII redaction, content moderation)
• Tool permission scopes (least privilege)
• Audit log of every tool call

Application Layer

• Human review for high-stakes outputs
• Clear AI disclosure to end users
• Feedback channels for users to report errors
• Continuous evaluation against measured outcomes

A Compliance Document Template

Most teams pursuing RMF alignment produce three documents:

flowchart TB
 Doc1[System Description<br/>scope, context, users] --> Pkg
 Doc2[Risk Assessment<br/>identified risks, controls] --> Pkg
 Doc3[Evaluation Results<br/>measurements, metrics] --> Pkg[Compliance Package]

These map cleanly to the GOVERN, MAP, MEASURE functions. MANAGE is operationalized in the controls themselves and in incident-response runbooks.

What Tools Help

By 2026 several open-source and commercial tools map directly to RMF controls:

• Garak (NVIDIA, open-source) — automated LLM vulnerability scanning
• Promptfoo — prompt evals with safety dimensions
• Inspect AI — UK AI Safety Institute's eval framework
• HiddenLayer and Robust Intelligence — commercial AI red-teaming and monitoring
• AWS Bedrock Guardrails / Azure Content Safety / Vertex AI Safety — cloud-native input/output guards

Common Gaps in Practice

What teams typically miss when first attempting RMF alignment:

• Eval cadence: one-time evaluations are not enough; the framework expects continuous measurement
• Incident reporting: most teams have no defined process for AI-specific incidents
• Provenance: data and model provenance is often undocumented
• Decommissioning: the framework includes managed retirement of models; rarely planned for

RMF and Other Frameworks

In 2026 the Generative AI Profile aligns reasonably with:

• ISO/IEC 42001 (AI management systems standard)
• EU AI Act high-risk requirements
• HIPAA Privacy Rule (when applied to PHI in AI)
• (when AI processes customer data)

A single internal compliance program can typically satisfy multiple frameworks.

What This Means for Builders

If you sell to enterprises or US federal customers in 2026, RMF alignment is increasingly a procurement requirement. The cost is real but bounded — typically a few months of focused work to set up, then ongoing measurement effort. Done well, the same investment supports EU AI Act, ISO 42001, and most enterprise risk reviews.

Sources

• NIST AI Risk Management Framework — https://www.nist.gov/itl/ai-risk-management-framework
• NIST AI 600-1 Generative AI Profile — https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
• ISO/IEC 42001 — https://www.iso.org
• Garak LLM scanner — https://github.com/NVIDIA/garak
• "Mapping AI risks to controls" CSA — https://cloudsecurityalliance.org