NIST AI RMF checklist
The NIST AI Risk Management Framework 1.0 groups its outcomes into four functions and 19 categories. Below, each category is written as a question you can answer about your own AI systems, with the evidence an assessor or a customer's security team is likely to ask you for.
Where to start if you have a month
The framework does not set an order. These six give you something to show quickly and make the rest easier.
- GOVERN 2Name an owner for every AI system. Every other step needs an owner to sign off.
- MAP 1Write one page per system on what it is for and who uses it.
- MAP 4List the third-party models, data and plugins inside each system.
- GOVERN 6Send your main SaaS vendors questions about the AI in their products.
- MEASURE 2Test your highest-risk system for prompt injection and data leakage.
- MANAGE 4Add AI to your incident plan, with who can switch a system off.
Govern
Who decides, who is accountable, and which rules apply to every AI system you run or buy.
| Category | The question | Evidence to keep |
|---|---|---|
| GOVERN 1 | Do you have written policies for finding, measuring and handling AI risk, and are they followed? | An AI policy with an owner and a review date, and a record of at least one decision made under it. |
| GOVERN 2 | Is it clear who is accountable for each AI system, and have those people been trained for it? | A named owner per system in your AI inventory, and training records for owners. |
| GOVERN 3 | Do the people who assess AI risk come from enough backgrounds to spot problems one group would miss? | The membership of your AI review group, covering at least security, legal or privacy, the business and the people affected. |
| GOVERN 4 | Do teams talk openly about AI risk, and can someone raise a concern without it being buried? | A route for raising AI concerns, and examples of concerns that were raised and answered. |
| GOVERN 5 | Do you collect feedback from people outside the build team, including users and those affected? | Feedback channels per system and a log of what came in and what changed. |
| GOVERN 6 | Do your rules cover AI you buy and the data and models you get from others? | Vendor questions on AI use, contract clauses on training on your data, and a list of third-party models in use. |
Map
What each system is for, where it runs, what it touches, and who could be hurt if it goes wrong.
| Category | The question | Evidence to keep |
|---|---|---|
| MAP 1 | For each system, is the purpose, the setting and the expected users written down? | One page per system: purpose, users, setting, and the decisions it informs or makes. |
| MAP 2 | Have you classified each system, for example by task, autonomy and the data it uses? | A risk tier or category recorded in the inventory, with the reason. |
| MAP 3 | Do you know what the system is meant to deliver and what it costs, compared with doing it another way? | Expected benefits, known limits and costs, compared with the current process or a baseline. |
| MAP 4 | Are risks mapped for every part of the system, including third-party models, data and plugins? | A component list or architecture diagram with risks recorded against each part. |
| MAP 5 | Have you described the likely impact on people, groups and the organisation if the system fails or is misused? | An impact assessment, a DPIA where personal data is involved, and how likely and how severe each impact is. |
Measure
How you test the risks you mapped, and how you keep testing once the system is live.
| Category | The question | Evidence to keep |
|---|---|---|
| MEASURE 1 | Have you chosen methods and metrics for the risks that matter most, and do you use them? | A test plan per system naming the metrics, and the latest results. |
| MEASURE 2 | Is each system tested for the trustworthy characteristics NIST lists: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed? MEASURE 2 also asks about environmental impact. | Test results per quality, including security tests such as prompt injection and data leakage. |
| MEASURE 3 | Can you track a risk over time, including new risks that show up in production? | Monitoring on live systems, and a risk register with dates and owners. |
| MEASURE 4 | Do you check whether your measurements work, using feedback from users and experts? | A review of your metrics, and changes made after feedback. |
Manage
What you do about each risk, and how you respond when something goes wrong.
| Category | The question | Evidence to keep |
|---|---|---|
| MANAGE 1 | Are the risks you found ranked, and has someone decided to fix, accept, transfer or avoid each one? | A risk register showing the decision, the owner and the date for each risk. |
| MANAGE 2 | Do you have plans to get the most benefit with the least harm, including a way to stop or roll back a system? | A tested way to switch off or roll back each system, and who can trigger it. |
| MANAGE 3 | Are risks from third-party models, data and services monitored and handled? | Vendor reviews on a schedule, and what you did when a vendor changed a model or a term. |
| MANAGE 4 | Are response, recovery and communication plans for AI incidents written down and checked regularly? | An AI section in your incident plan, and the date it was last exercised. |
Using generative AI?
NIST published a Generative AI Profile in July 2024 (NIST AI 600-1). It keeps the same four functions and adds actions for risks specific to generative models, such as confabulation, data privacy, information security and dangerous, violent or hateful content. Use the checklist above first, then the profile for any system built on a large language model.
Also working to the UK AI Cyber Security Code?
The UK code's 13 principles call for much of the same evidence. Our UK checklist has a table showing which NIST AI RMF categories the same evidence covers.
Want a score against the framework?
Our free assessment asks the NIST AI RMF 1.0 questions for you and gives a short report with the gaps that matter most and what to do first. The same assessment also covers the NIST Generative AI Profile, the EU AI Act and the UK AI Cyber Security Code of Practice.
Sources
- NIST AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1, January 2023)
- AI RMF Core: functions and categories (NIST AI Resource Center)
- NIST AI 600-1, Generative Artificial Intelligence Profile (July 2024)
Checked against the sources on 15 November 2026.
The questions are our plain-English reading of each category. Free to use and adapt inside your organisation. Not legal advice.