The NIST AI Risk Management Framework is voluntary guidance for managing AI risk. If you are the first security hire at a company that already uses AI, begin with one register of AI uses, one named owner for each and one page that says which harms you care about most. The framework's four functions then give you a structure.
What should you read first?
NIST's AI RMF page explains the purpose and links to the framework document, the playbook and related profiles. The framework was released on 26 January 2023 after a consensus driven process, and NIST describes it as intended for voluntary use. It is a map for your thinking, not a checklist you can fail.
Read the short overview first. Then read only the playbook sections for the functions you are working on this quarter.
What are the four functions?
- Govern: who is accountable, what policies exist, what risk the company will accept.
- Map: what each AI system does, in what setting, and who could be affected.
- Measure: how you test and track whether the system behaves as intended.
- Manage: how you respond to risks, change systems and communicate.
Govern runs through everything else. Without owners and a risk appetite, the rest becomes paperwork.
What do you do in the first month?
Talk to people. Ask engineering, product, support, marketing and finance what AI tools and features they use or build. You will find some of it in code and more of it in browsers and expense reports. For each item write down: what it is, who owns it, what data it touches, who is affected by its output, and the vendor if any.
A constructed finding: the sales team connects a meeting recorder with AI summaries to the company calendar. It stores call transcripts with a vendor nobody reviewed. That is a supplier risk as much as an AI one, and you can handle it with your normal vendor process.
How do you pick what to work on?
Rank uses by impact and exposure. A customer facing assistant with access to account data outranks an internal tool that rewrites meeting notes. For the top few, ask four questions: what is the worst plausible failure, how would we notice, who can switch it off, and what evidence do we keep? Those questions are Map, Measure and Manage in plain words.
Which AI specific risks should you know?
Security teams should be familiar with prompt injection, leakage of sensitive data into prompts or outputs, agents with excessive permissions, poisoned or untrusted data sources and weak model or dependency provenance. The OWASP project on LLM applications is a good practical companion to the NIST framework.
What not to do
Do not announce a program before you have an inventory. Do not copy a long policy from a template that nobody will follow. Do not wait for perfect measures; start with simple tests and improve them.
What to do
- Read NIST's overview and skim the playbook for Govern.
- Interview teams and build the AI register.
- Assign an owner to each entry.
- Rank the top five by impact and write the four questions for each.
- Add AI vendors to your supplier review.
- Write a one page AI use policy with approved tools and forbidden data.
- Report progress to leadership every quarter in plain language.
Quick check
Can a company be certified against the NIST AI RMF?
No. NIST publishes it as voluntary guidance. You can describe how your controls align to it, but there is no official certificate.
What is the first deliverable?
A register of AI uses with an owner for each. It is the base for every other step.
Sources
Last reviewed: 18 October 2026.
