AI risk management is the practice of identifying, measuring and reducing the harms an AI system can cause, to people, to the organization and to society. The NIST AI Risk Management Framework organizes it into four functions: Govern, Map, Measure and Manage. It is voluntary, and applies to any organization building or using AI.
What is the NIST AI RMF?
NIST released the framework on 26 January 2023 for voluntary use, to "improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems". A companion playbook, a roadmap and a crosswalk accompany it. On 26 July 2024 NIST released NIST AI 600-1, a Generative AI Profile, which identifies risks unique to generative AI and proposes actions to manage them. NIST's page also notes the AI RMF 1.0 is being revised, and that in April 2026 it released a concept note for a profile on trustworthy AI in critical infrastructure. Check the page for the current version before you build a program around it.
What do the four functions mean?
Govern sets policy, roles and accountability for AI across the organization. Map establishes context: what the system is for, who it affects, and what could go wrong. Measure uses tests and metrics to assess those risks. Manage prioritizes and acts on them, including monitoring after launch.
What are the benefits?
Clear ownership: someone is accountable for each AI system, so decisions do not drift. Earlier discovery of problems, since mapping and testing happen before launch. A common language for engineering, legal, security and product teams. And evidence for customers and auditors that you manage AI deliberately.
What are the challenges?
Inventory comes first and is harder than it sounds. Teams adopt AI tools without telling security, and models arrive embedded in vendor products. You cannot manage risk in systems you cannot list.
Measurement is difficult. Model behavior is probabilistic, so a test that passes once does not prove the next output is safe. Plan for ongoing evaluation, not a single sign off.
Third party models add dependencies you do not control. Ask suppliers what data they train on, how they handle your inputs and how they notify you of changes.
Security risks overlap with ordinary application security, and are listed in resources such as the OWASP Top 10 for LLM applications. Risk management should connect to those controls, not sit beside them.
Pace is a challenge too. Frameworks change, as NIST's own revision notice shows, so a program tied to one document needs a review date.
What does a first program look like?
Keep it small. Start with a list of AI systems and who owns each. Add a short intake form asking what data the system sees, what it can do, and who is affected. Rate each system as low, medium or high impact, with extra review for high. Define testing before launch and monitoring afterward. Decide who can approve exceptions.
A constructed example
A company finds seven AI features: a support chatbot, a code assistant, a resume screening tool, and four small internal automations. It rates the resume tool high impact because it affects hiring decisions about people, and requires human review of rejections plus periodic bias testing. The chatbot is medium: limited to public knowledge with no account access. The internal automations are low. Review effort follows impact.
How to check or apply this
- Build an inventory of AI systems, including those inside vendor products.
- Assign an owner and an impact rating to each.
- Write an intake and review process for new AI use.
- Test before launch against your risks, and repeat on a schedule.
- Monitor behavior, incidents and complaints after launch.
- Review the program when the framework or your systems change.
Safeguard's MCP Guardrail, agent identity and policy features address tool access and permissions for AI agents, and its compliance automation can organize evidence.
Quick check
Is the NIST AI RMF mandatory?
No. NIST describes it as intended for voluntary use.
What is NIST AI 600-1?
The Generative AI Profile released on 26 July 2024, which helps organizations identify risks unique to generative AI and choose actions to manage them.
Sources
Last reviewed: 18 October 2026.
