The NIST AI Risk Management Framework defines four functions: Govern, Map, Measure, Manage. Govern sets organisational culture and accountability; Map identifies the AI system's context and risks; Measure quantifies and tracks them; Manage takes action and monitors. The NIST Generative AI Profile (NIST AI 600-1) layers GenAI-specific risks on top, CBRN information, confabulation, harmful bias, value-chain risks, environmental impact, intellectual-property leakage, obscene/violent content, dangerous capabilities, and maps each to controls. It is voluntary but widely cited as the US baseline for AI governance.
The NIST AI Risk Management Framework is the United States government's voluntary framework for managing AI risk. It is structured around four functions that together describe the lifecycle of risk management.
- **Govern**, set the organisational culture, policies, accountability, and roles that make AI risk management possible. This is the foundation that the other three functions rest on.
- **Map**, understand the context. What is the AI system? Who uses it? What does it do? What risks does the deployment context introduce?
- **Measure**, quantify and track the risks identified in Map. Use tests, evaluations, metrics, and red-team exercises.
- **Manage**, take action on the measured risks. Prioritize, mitigate, monitor, and continuously update.
The NIST Generative AI Profile (NIST AI 600-1, released 2024) extends the framework with risks specific to generative AI: dangerous CBRN information, confabulation (hallucination), harmful bias, intellectual-property leakage, value-chain risks (third-party model and data dependencies), environmental impact, obscene or violent content, and dangerous capabilities. Each risk is mapped to controls under the four functions.
The framework is voluntary, but it is widely cited as the US baseline for AI governance and forms the basis for many enterprise AI policies in 2025-2026.
**The NIST AI Risk Management Framework defines four functions that together describe the AI risk-management lifecycle.**
- **Govern** sets organisational culture, accountability, roles, policies, and risk tolerances. It is the foundation that the other three functions depend on.
- **Map** identifies the AI system's context: purpose, users, stakeholders, data sources, and the risks introduced by the deployment environment.
- **Measure** quantifies the mapped risks through tests, evaluations, metrics, red-team exercises, and ongoing monitoring.
- **Manage** takes action on the measured risks, prioritization, mitigation, response, and continuous review.
The **NIST Generative AI Profile (NIST AI 600-1, July 2024)** layers GenAI-specific risk categories on top:
- **CBRN information**, chemical, biological, radiological, nuclear uplift
- **Confabulation**, fabricated facts (hallucination)
- **Harmful bias and homogenization** of viewpoints
- **Value-chain and intellectual-property risks**, third-party model dependencies, training-data IP
- **Information integrity**, deepfakes and synthetic media
- **Information security**, model weights leakage, prompt injection
- **Obscene and abusive content**
- **Dangerous capabilities**, autonomy and self-replication concerns
- **Environmental impact**, training and inference energy
Each risk maps to recommended controls under Govern, Map, Measure, and Manage. The framework is voluntary, but it is the de facto US baseline. Enterprise AI policies in 2025-2026 commonly reference it, and federal procurement guidance increasingly expects alignment.
Quick reference
TL;DR
Govern, Map, Measure, Manage are the four NIST AI RMF functions; the NIST Generative AI Profile (AI 600-1) layers GenAI-specific risk categories like CBRN, confabulation, and IP leakage onto them.
Picture running a kitchen. You set the rules of the kitchen and who is responsible for what (Govern). You walk through the menu and figure out which dishes have allergens, sharp knives, or open flames (Map). You weigh ingredients, time cooks, and track complaints (Measure). You actually act, change the recipe, add a sign, retrain a cook (Manage). Now imagine you start serving experimental new dishes nobody has cooked before. You add an extra checklist for those: new allergens to watch for, new ways things can go wrong. That extra checklist is the GenAI Profile sitting on top of the basic four-function kitchen plan.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Deployment outside the United States where the EU AI Act, UK AI policy, or other regional regulation is the binding requirement The AI RMF is voluntary US guidance; binding regulation in other jurisdictions has its own taxonomy and obligations
Internal research workload with no users, no deployment surface, and no third-party data The framework adds documentation overhead that may not be justified for purely internal experimentation
Highly regulated domain where a specific sectoral framework (HIPAA, PCI-DSS, SR 11-7) takes precedence AI RMF is general guidance; sectoral frameworks have specific obligations that override or supplement it
Quick scan · 2–4 min
Everything important, quickly.
5 min: four functions, role of each, GenAI Profile overlay, key new risk categories, voluntary status, contrast with binding regulation.
Real-world examples
Real products, models, and research that use this idea.
1Enterprise procurement teams in 2025-2026 commonly require vendors to document AI RMF alignment, often citing AI 600-1 control mappings
2US federal agencies use the AI RMF as the baseline for internal AI governance under OMB M-24-10 and successor memoranda
3Foundation-model providers (Anthropic, OpenAI, Google, Meta) publish model cards and system cards that map to NIST risk categories
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does the AI RMF relate to the EU AI Act?
Sign in or go Pro to see the approach for this follow-up.
QWhich GenAI risk category from AI 600-1 is most underweighted in current production stacks?
Sign in or go Pro to see the approach for this follow-up.
QHow would you operationalize the Measure function for a 2026 frontier-model deployment?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Treating the NIST AI RMF as a security checklist rather than a governance framework, and missing that the GenAI Profile is the layer where LLM-specific risks like confabulation and IP leakage appear.