meilynx

Reference · AI risk frameworks

NIST AI 600-1: The Generative AI Profile, Its 12 Risks and the Evidence an Assessor Asks For

The AI RMF companion for generative AI, read for the evidence it calls for

Last reviewed November 15, 2026

At a glance

Released
26 July 2024, by NIST
Legal effect
Voluntary; no certification
Contents
12 risk categories, 211 suggested actions
Parent framework
AI RMF 1.0 (NIST AI 100-1), under revision

NIST AI 600-1, the Generative AI Profile, is a companion to the NIST AI Risk Management Framework (AI RMF 1.0). NIST released it on 26 July 2024, and like the framework it is intended for voluntary use.NIST AI RMF

The profile names twelve risks that generative AI creates or makes worse, and sets out 211 suggested actions under the framework's four functions: Govern, Map, Measure and Manage.NIST AI 600-1 · Sections 2 and 3 An organization chooses the actions that fit its role; the profile notes that an action relevant to a developer may not be relevant to a deployer.NIST AI 600-1 · Section 3

A profile applies the AI RMF to one technology

In NIST's terms, a profile implements the AI RMF's functions, categories and subcategories for a specific setting or technology. AI 600-1 is cross-sectoral: it covers generative AI wherever it is built or used, including the use of large language models, cloud-based services and acquisition.NIST AI 600-1 · Section 1

Its suggested actions concentrate on four considerations from NIST's Generative AI Public Working Group: governance, content provenance, pre-deployment testing and incident disclosure. It addresses a subset of the AI RMF's subcategories and leaves out speculative risks of future systems.NIST AI 600-1 · Sections 1 to 3 Each action carries an identifier tied to its function and subcategory, so MG-4.3-002 is the second suggested action under MANAGE 4.3.NIST AI 600-1 · Section 3

Measure holds the most actions, and Measure and Manage together hold more than half.

Bars

Suggested actions in NIST AI 600-1, by AI RMF function

Actions 050100Govern57Map39Measure72Manage43
Counted from the action tables in Section 3 of NIST AI 600-1, 211 actions in all.

Twelve risks that generative AI creates or makes worse

NIST limits the list to risks with an existing evidence base and labels each by its outcome, object or source.NIST AI 600-1 · Section 2 Several reach any organization that deploys a language model; others weigh mostly on the developers of the model.

  • CBRN information or capabilities: easier access to, or synthesis of, information and design capabilities for chemical, biological, radiological or nuclear weapons.NIST AI 600-1 · §2.1
  • Confabulation: confidently stated but false content, commonly called hallucination, that can mislead the people who rely on it.NIST AI 600-1 · §2.2
  • Dangerous, violent or hateful content: easier production of inciting, radicalizing or threatening content, and of recommendations to self-harm or to act illegally.NIST AI 600-1 · §2.3
  • Data privacy: leakage and unauthorized use, disclosure or de-anonymization of personal or sensitive data, including data memorized in training.NIST AI 600-1 · §2.4
  • Environmental impacts: the compute and energy that training and operating a model consume, and their effect on ecosystems.NIST AI 600-1 · §2.5
  • Harmful bias and homogenization: amplified historical and systemic bias, performance gaps between groups or languages, and outputs that converge into undesired uniformity.NIST AI 600-1 · §2.6
  • Human-AI configuration: how people and the system interact, including automation bias, over-reliance, anthropomorphizing and emotional entanglement.NIST AI 600-1 · §2.7
  • Information integrity: a lower barrier to producing content that blurs fact and opinion, and to running disinformation campaigns at scale.NIST AI 600-1 · §2.8
  • Information security: lower barriers to offensive cyber operations, and a larger attack surface for the system itself through direct and indirect prompt injection and data poisoning.NIST AI 600-1 · §2.9
  • Intellectual property: easier replication of copyrighted, trademarked or licensed content, and exposure of trade secrets.NIST AI 600-1 · §2.10
  • Obscene, degrading or abusive content: easier production of abusive imagery, including synthetic child sexual abuse material and non-consensual intimate images.NIST AI 600-1 · §2.11
  • Value chain and component integration: third-party models, data and components that are integrated without transparency, improperly obtained or poorly vetted.NIST AI 600-1 · §2.12

Measure and Manage are where an assessor asks for production records

Most Govern and Map actions are answered with documents: policies, risk tiers, inventories and contracts. A set of Measure and Manage actions asks what the system did once it was live, and that is shown only by records kept from its traffic. The actions below are NIST's; the question attached to each is our reading of what an assessor asks for.

  • Monitoring after deployment. MG-4.1-002 asks for processes to monitor generative AI systems after deployment, particularly for confabulation, CBRN and cyber risks. The assessor asks for the monitoring record and the findings it produced over the period.NIST AI 600-1 · MG-4.1-002
  • Errors, near misses and incidents. MANAGE 4.3 asks for policies to record and track reported errors, near misses and negative impacts, after-action reviews of incidents, and reporting where law requires it. The assessor asks for the incident log, with each entry traced to the events behind it.NIST AI 600-1 · MG-4.3-001 to 003
  • Escalation and deactivation. MANAGE 2.4 asks for criteria that warrant deactivating a system and for escalation to the risk authority when they are met. The assessor asks which events met the criteria and what followed.NIST AI 600-1 · MG-2.4-002, MG-2.4-004
  • Content provenance. MS-2.8-003 suggests documenting each instance where content is generated, modified or shared, to give "a tamper-proof history of the content". The assessor asks for that history and how its integrity is checked.NIST AI 600-1 · MS-2.8-003
  • Input and output filters. MG-3.2-005 suggests filters, rule-based or model-based, that flag problematic inputs and outputs. The assessor asks for the filters in force and what they flagged.NIST AI 600-1 · MG-3.2-005
  • Security measures in operation. MS-2.7-004 names metrics such as the number of unauthorized access attempts, and MS-2.7-009 asks for regular checks that security measures remain effective. The assessor asks for the counts and the checks; the control-level detail is in NIST SP 800-53 and AI systems.NIST AI 600-1 · MS-2.7-004, MS-2.7-009
  • Third-party models. MANAGE 3.1 asks for the organization's risk tolerances and controls, input and output filtering among them, to be applied to third-party generative AI, and for review of third-party system and model cards. The assessor asks which external models the system actually called, and under which controls.NIST AI 600-1 · MG-3.1-001, MG-3.1-005

The profile gives structure to rules financial services already follow

SR 26-2, the model risk guidance the Federal Reserve, the OCC and the FDIC issued on 17 April 2026, places generative and agentic AI models outside its scope and leaves their governance to the bank's own risk management practices.Revised Guidance · footnote 3 The profile is a ready structure for that governance, and the Treasury's Financial Services AI Risk Management Framework, released on 19 February 2026, adapts the same NIST framework to financial services.Treasury SR 26-2 and the generative AI carve-out cover the guidance in full.

FINRA Regulatory Notice 24-09 creates no new requirements and applies FINRA's rules to generative AI as to any other technology. A member firm that uses generative AI in its supervisory system should address technology governance in its procedures, "including model risk management, data privacy and integrity, reliability and accuracy of the AI model."Notice 24-09 The profile's actions on confabulation, data privacy and information security cover the same ground. More in FINRA Regulatory Notice 24-09.

ISO/IEC 42001 specifies an AI management system that an organization can certify, and NIST publishes a crosswalk from the AI RMF to the standard.NIST crosswalk An organization that runs a 42001 management system can take the profile's actions as generative AI controls inside it, and the same Measure and Manage records serve the standard's clauses on operation and performance evaluation.

The AI RMF is under revision, and the 2024 profile is still the current text

America's AI Action Plan, released by the White House on 23 July 2025, recommends that NIST revise the AI RMF to remove references to misinformation, diversity, equity and inclusion, and climate change.White HouseAI Action Plan · p. 4 NIST states that AI RMF 1.0 is being revised as part of the plan.NIST AI RMF

On 7 April 2026 NIST released a concept note for an AI RMF profile on trustworthy AI in critical infrastructure.NIST AI RMF As of this page's review date, NIST has published no revised framework text or Generative AI Profile. AI RMF 1.0 and AI 600-1 remain the current documents.

Read against the profile, the subjects the plan names correspond most closely to three of its twelve categories: information integrity, environmental impacts, and harmful bias and homogenization.

Control mapping

What a reviewer expects to be able to see.

ObligationWhat the system must doEvidence a reviewer expects
Inventory generative AI systems (GOVERN 1.6)Enumerate generative AI systems, with the underlying models, versions and access modes, and record any exemptions in policyAn inventory reconciled against the models the systems actually call
Evaluate security and resilience (MEASURE 2.7)Track security metrics such as unauthorized access attempts, red-team against prompt injection, and check that measures remain effectiveAccess-attempt counts, red-team results and dated effectiveness checks
Examine transparency and accountability (MEASURE 2.8)Document each instance where content is generated, modified or shared, with a tamper-proof historyAn integrity-checked record of each generation and its policy outcome
Monitor third-party generative AI (MANAGE 3.1)Apply risk tolerances and controls, including input and output filtering, to third-party modelsPer-provider usage records and the controls applied to each call
Monitor pre-trained models and filter content (MANAGE 3.2)Run content filters on inputs and outputs and monitor generated content in real timeFilter configuration in force and the events it flagged
Monitor after deployment (MANAGE 4.1)Implement post-deployment monitoring, particularly for confabulation, CBRN and cyber risksThe monitoring record and its findings across the review period
Track and communicate incidents (MANAGE 4.3)Record errors, near misses and negative impacts, review incidents after the fact, and report where law requiresAn incident log traced to the underlying events, with after-action reviews

Key dates

Timeline, drawn to scale

The AI RMF and its Generative AI Profile, from release to revision

PublishedThe profile and the revision
20242025202626 JAN 2023AI RMF 1.026 JUL 2024AI 600-123 JUL 2025Revision recommended7 APR 2026Concept note
Dates from the NIST AI RMF page and the White House. Status as of November 15, 2026.

Sources cited

Common gaps

Where a generative AI program most often comes up short against the profile's Measure and Manage actions.

  • An inventory built from a survey. Generative AI embedded in applications, and models reached through a provider's API, are the systems a survey misses. GV-1.6-002 asks for any exemption to be defined in policy.NIST AI 600-1 · GV-1.6-002
  • Monitoring that keeps no record. A dashboard that shows current alerts and keeps no history cannot show what post-deployment monitoring found over the period under review (MG-4.1-002).
  • An incident log with no link to the events. When an entry cannot be traced to the requests and responses behind it, the after-action review that MG-4.3-001 calls for has no facts to work from.
  • Third-party models outside the controls. A provider-hosted or fine-tuned model called without the input and output filters applied to in-house systems falls short of MG-3.1-001, which applies the same risk tolerances to third-party resources.
  • A provenance history that can be edited. Application logs an administrator can edit fall short of the tamper-proof history MS-2.8-003 describes, which needs an integrity check over the record.

Last reviewed November 15, 2026. This reference summarises publicly available regulatory guidance and is provided for general information. It is not legal advice. Obligations depend on an institution's charter, registration status, size, and activities. Verify against the primary sources cited above and consult counsel before relying on any summary here.

Regulatory updates

When a regulator changes what an AI examination asks for, hear about it first.

Short notes on SR 26-2, NYDFS 500, FINRA, the NAIC bulletin, the EU AI Act, and the employment-AI statutes, plus what we ship. A few emails a month.