Compliance Guide - NIST AI Risk Management Framework¶

Overview¶

The NIST AI Risk Management Framework (AI RMF 1.0, released January 2023) is a voluntary, sector-agnostic framework for managing risks in the design, development, deployment, and use of AI systems. It is the US counterpart to the EU's regulatory approach: nobody fines you for ignoring it, but it has become the default vocabulary for AI governance in US enterprise procurement, and federal agency guidance increasingly references it.

Its practical value is that it is implementable. It gives you four functions, a set of trustworthiness characteristics to measure against, and a companion playbook of concrete actions, without prescribing a specific technology or control set.

📖 NIST AI Risk Management Framework - the framework hub, playbook, and companion resources 📖 AI RMF 1.0 (NIST AI 100-1) - the framework document 📖 Generative AI Profile (NIST AI 600-1) - GenAI-specific risks and suggested actions

The four functions¶

flowchart TB
  GOV[GOVERN<br/>culture, policy, accountability<br/>cuts across everything]

  subgraph cycle[Applied per AI system, iteratively]
    MAP[MAP<br/>context, purpose,<br/>risks identified]
    MEAS[MEASURE<br/>analyze, benchmark,<br/>track metrics]
    MAN[MANAGE<br/>prioritize, respond,<br/>monitor, recover]
  end

  GOV --> MAP
  MAP --> MEAS
  MEAS --> MAN
  MAN --> MAP
  GOV -.governs.-> MEAS
  GOV -.governs.-> MAN

Govern is continuous and organizational. Map, Measure, and Manage repeat per system and per material change.

Govern¶

The organizational layer: policies, roles, accountability, culture, and workforce competence.

Typical evidence: an AI policy, an AI use inventory, a named accountable owner per system, a review board or approval gate, third-party and supply chain requirements, incident response procedures that cover AI failure modes, and documented staff training.

If you only do one thing from AI RMF, do the inventory. Almost every governance gap traces back to an organization not knowing which AI systems it operates.

Map¶

Establish context and identify risks before building. What is the system for, who is affected, what does success mean, what happens when it fails, and what are the alternatives to building it at all.

Typical evidence: intended purpose and out-of-scope uses, affected stakeholder analysis, data provenance and lineage, documented assumptions and limitations, third-party component inventory, and a recorded decision that AI is an appropriate solution.

Measure¶

Quantify. Test, evaluate, verify, and validate against the trustworthiness characteristics, then keep measuring in production.

Typical evidence: an eval suite with baselines, fairness and bias metrics across relevant subgroups, robustness and adversarial testing results, explainability assessments, human factors testing, drift monitoring, and a documented method for what is measured, how often, and by whom.

This is where the framework meets engineering practice most directly. See Evals for LLMs and LLM red teaming.

Manage¶

Act on what you measured: prioritize risks, allocate resources, apply treatments, monitor, and plan for recovery and decommissioning.

Typical evidence: a risk register with treatment decisions and residual risk sign-off, deployment gates tied to measured thresholds, monitoring and alerting, an incident response runbook covering model failures, a rollback and model version pinning strategy, and a decommissioning plan.

Trustworthiness characteristics¶

The framework defines what "trustworthy" means, so that Measure has targets.

These trade off against one another. Explainability may cost accuracy; privacy techniques may cost utility. The framework expects you to document the trade-off rather than pretend it does not exist.

The Generative AI Profile¶

NIST AI 600-1 is a companion that applies the four functions to generative AI specifically. It identifies risks that are unique to or amplified by GenAI, including confabulation (hallucination), dangerous or violent content, data privacy leakage, environmental impact, harmful bias, human-AI configuration problems, information integrity, information security, intellectual property, obscene content, and value chain integration risk.

For each, it proposes suggested actions mapped to Govern, Map, Measure, and Manage. In practice this is the most directly usable NIST artifact for a team building on foundation models, because it names the failure modes engineers actually encounter.

Companion resources worth knowing: the AI RMF Playbook (concrete suggested actions per subcategory), NIST AI 100-2 on adversarial machine learning taxonomy, and the Secure Software Development Practices for Generative AI community profile.

Implementing it¶

A realistic sequence for an organization starting from nothing.

Phase 1 - Govern the basics (weeks 1 to 6) 1. Build the AI system inventory. Include shadow AI: the team using a chat product with company data is in scope. 2. Write a one-page AI policy: what requires review, what is prohibited, who approves. 3. Name an accountable owner per system. 4. Define risk tiers using your own criteria, and align them with EU AI Act classification if you have EU exposure. One classification exercise, two outputs.

Phase 2 - Map and Measure the highest-tier systems (weeks 6 to 16) 5. For each high-tier system: document purpose, stakeholders, data provenance, limitations, and out-of-scope uses. 6. Stand up an eval suite with baselines. Include safety and injection tests, not just accuracy. 7. Run bias and robustness testing; record results and the subgroups examined.

Phase 3 - Manage and operationalize (ongoing) 8. Risk register with treatments and residual risk acceptance signed by the accountable owner. 9. Deploy gates: no release if the safety suite regresses. 10. Production monitoring for drift, cost, refusal rate, and incident signals. 11. Incident runbook and a decommissioning plan. 12. Annual review, and re-assessment whenever purpose, model, or data changes materially.

Cloud provider mapping¶

Relationship to other frameworks¶

Documentation links¶

📖 NIST AI RMF hub - framework, playbook, crosswalks, and profiles 📖 AI RMF 1.0 (AI 100-1) - the framework document 📖 Generative AI Profile (AI 600-1) - GenAI risks and suggested actions 📖 Adversarial Machine Learning taxonomy (AI 100-2) - attack and mitigation taxonomy 📖 NIST Trustworthy and Responsible AI Resource Center - playbook and crosswalk tooling