What is the NIST AI RMF?
NIST released AI RMF 1.0 (NIST AI 100-1) on 26 January 2023. It does two things. It defines what trustworthy AI looks like, as seven characteristics. And it organizes the work of managing AI risk into four functions, broken into categories and subcategories, with a companion Playbook of suggested actions. It is voluntary, sector-neutral and technology-neutral, which is both why it spread and why it needs translation before an engineering team can use it.
The seven trustworthiness characteristics:
- Valid and reliable
- Safe
- Secure and resilient
- Accountable and transparent
- Explainable and interpretable
- Privacy-enhanced
- Fair, with harmful bias managed
They trade off against each other. A more explainable model can be less accurate; stronger privacy can make bias harder to measure. The framework does not resolve those trade-offs. It asks you to make them deliberately and write down why.
Govern, Map, Measure, Manage
| Function | Purpose | Evidence that satisfies it |
|---|---|---|
| Govern | Cross-cutting: policies, accountability, risk tolerance, workforce, third-party AI, a culture of raising risks | AI policy, RACI, risk appetite statement, vendor AI terms, training records |
| Map | Establish context for each system: purpose, users, impacts, assumptions, categorization | Use case intake, impact assessment, risk tier, documented intended and prohibited uses |
| Measure | Analyze and track the risks identified: test, evaluate, verify, validate, monitor | Evaluation suites, bias and robustness tests, red-team results, production metrics |
| Manage | Prioritize and act: treat, transfer, accept or avoid risks; respond to incidents; decommission | Risk treatment decisions, incident runbooks, change records, retirement records |
Govern is cross-cutting and comes first for a reason. Map, Measure and Manage are performed per system, and without a policy that says which systems need how much of each, every team decides for itself. The result is the most common failure we see: a beautifully documented risk assessment for one flagship model and nothing for the forty internal tools built on the same API.
The practical shortcut is to tie depth to risk tier. A low-risk internal summarizer gets a one-page Map and a standard evaluation. A system that scores job applicants gets the full treatment, external testing and a named owner for every Manage decision. The framework allows exactly this; it says organizations should apply it in proportion to risk.
The Generative AI Profile (NIST AI 600-1)
On 26 July 2024 NIST published a profile for generative AI. It names twelve risks that generative systems create or make worse, and maps more than 200 suggested actions to the four functions:
- CBRN information or capabilities
- Confabulation
- Dangerous, violent or hateful content
- Data privacy
- Environmental impacts
- Harmful bias and homogenization
- Human-AI configuration
- Information integrity
- Information security
- Intellectual property
- Obscene, degrading or abusive content
- Value chain and component integration
For most enterprises the useful ones are confabulation, information security (prompt injection, data leakage), intellectual property, human-AI configuration (over-reliance and automation bias) and value chain risk, which is the risk inherited from the model provider and every component you did not build.
Where the framework stands in 2026
The core is still 1.0. NIST has said it is revising the framework in response to the White House AI Action Plan of July 2025, but it has not published a revised core. What it has published this year sits around the edges:
- 17 February 2026: the AI Agent Standards Initiative, led by NIST's Center for AI Standards and Innovation (CAISI), covering agent standards, open agent protocols and research on agent security and identity, with requests for input on agent security and on agent identity and authorization.
- 7 April 2026: a concept note for an AI RMF profile on trustworthy AI in critical infrastructure.
- 29 July 2026: an initial public draft of guidance and templates for public-facing AI documentation (NIST AI 300-1).
The direction is sector and technology profiles layered on a stable core. That is good news for anyone who built a program on 1.0: the investment carries forward.
Applying it to AI agents
The framework was written when most AI was a model behind an application. Agents change the unit of analysis. An agent's risk is not its model; it is the model times the tools, data and permissions it can reach, at the moment it acts. The four functions still hold. What fills them changes.
| Function | What changes for agents |
|---|---|
| Govern | Treat each agent as an accountable identity with an owner, scoped permissions and an approval record. Add agent-specific terms to vendor contracts: logging, kill switch, data use. |
| Map | Map tools and data an agent can reach, not only its model. The impact of an agent is the union of its permissions. |
| Measure | Move from release-time evaluation to continuous measurement: task success, policy violations, tool-call anomalies, cost per task, escalation rate. |
| Manage | Enforce at runtime: gateway allow lists, rate limits, human approval for irreversible actions, and a tested way to stop an agent mid-task. |
Measure is where programs break first. A release-time evaluation of an agent that selects its own tools is a snapshot of one path through a large space. Continuous measurement in production, wired to Manage actions that can actually stop the agent, is the agent-era version of the same function. This is also the main reason AI governance platforms spent 2026 adding runtime enforcement; see what an AI governance platform does.
A 90-day starting plan
- Weeks 1 to 3, Govern: publish an AI policy, a risk tiering rule and a RACI. Name the executive who accepts residual risk.
- Weeks 2 to 6, Map: build the AI inventory and tier every system. Expect to find more than you thought.
- Weeks 5 to 10, Measure: standard evaluation suites per tier; full testing on the top five systems by risk.
- Weeks 8 to 13, Manage: treatment decisions on record, an AI incident runbook, and a tested way to roll back or switch off each high-tier system.
When you need external assurance, the same records feed an ISO 42001 certification. If your AI reaches EU users, map the work to the EU AI Act obligations as well.
Related questions
Frequently asked questions
What is the NIST AI RMF?
The NIST AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1) is a voluntary framework published by the US National Institute of Standards and Technology on 26 January 2023 to help organizations manage risks from designing, developing, deploying and using AI. It defines characteristics of trustworthy AI and organizes risk work into four functions: Govern, Map, Measure and Manage. A companion Playbook suggests actions for each.
Is the NIST AI RMF mandatory?
No. It is voluntary for private companies. It becomes effectively required when a customer, a federal contract, a regulator or a state law points to it. Several US state AI laws and procurement rules reference it or treat recognized frameworks such as the NIST AI RMF and ISO/IEC 42001 favorably, so check the rules for your sector and states.
What is NIST AI 600-1?
NIST AI 600-1 is the Generative AI Profile of the AI RMF, published on 26 July 2024. It describes twelve risks that are unique to or made worse by generative AI, from confabulation and information integrity to intellectual property and value-chain risks, and lists more than 200 suggested actions mapped to the four RMF functions.
Is there a NIST AI RMF 2.0?
Not as of November 2026. NIST has said AI RMF 1.0 is being revised in response to the White House AI Action Plan of July 2025, but it has not published a revised core. Its 2026 work has come as add-ons: an AI Agent Standards Initiative led by its Center for AI Standards and Innovation (February 2026), a concept note for a critical infrastructure profile (April 2026) and draft guidance on public-facing AI documentation, NIST AI 300-1 (July 2026).
NIST AI RMF or ISO 42001?
Use the NIST AI RMF to structure the risk work quickly and ISO/IEC 42001 when you need a certificate customers and auditors recognize. They overlap heavily, and NIST publishes crosswalks between the RMF and ISO AI standards. A common path is to build the program on the RMF functions and certify against ISO 42001 once the operating records exist.
How long does it take to implement the NIST AI RMF?
A first usable version, meaning a governance policy, an inventory, a risk tiering method and Map and Measure work on the highest-risk systems, typically takes one to two quarters. Full coverage is open-ended because the framework is meant to run continuously. Do not try to answer every subcategory for every system; scale depth with risk tier.
Sources
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, 26 January 2023; NIST, Generative Artificial Intelligence Profile, NIST AI 600-1, 26 July 2024.
- NIST, Announcing the AI Agent Standards Initiative (17 February 2026).
- Agentic Rail, NIST AI RMF in 2026: still 1.0, and the agentic evidence gap (October 2026), for the 2026 concept note and NIST AI 300-1 draft dates.