Product team testing a smart-speaker-style device with microphones in an anechoic chamber

Cyber Resilience Act

NIST AI 600-1: the Generative AI Profile, explained

Last updated:

NIST AI 600-1, the Generative AI Profile, is a voluntary companion to the NIST AI Risk Management Framework that NIST published on 26 July 2024 at the direction of a now-revoked executive order, naming 12 risks generative AI creates or makes worse and mapping 212 suggested actions to the base framework's four functions. It does not regulate anyone and was never mandatory outside federal agencies applying it under now-superseded OMB guidance — but it remains the most detailed public reference for what, specifically, generative AI changes about AI risk, and it is a genuinely useful checklist for a device or firmware maker scoping risk in a product that embeds or calls a generative model.

Key takeaways

  • NIST published the AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1) on 26 January 2023 as a voluntary, four-function structure — Govern, Map, Measure, Manage — for managing AI risk generally.
  • The Profile's §2 names 12 risks "unique to or exacerbated by" generative AI.
  • The final text holds 212 suggested actions.
  • When an enterprise or government buyer uses the Profile, its procurement and third-party actions turn into questions for the supplier.

What is NIST AI 600-1, and how does it relate to the base AI RMF?

NIST published the AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1) on 26 January 2023 as a voluntary, four-function structure — Govern, Map, Measure, Manage — for managing AI risk generally. Executive Order 14110, issued 30 October 2023, directed the Secretary of Commerce, acting through NIST, to produce a companion resource for generative AI; NIST AI 600-1 is that resource, released 26 July 2024 per NIST's own AI RMF page.

The published Profile defines a profile as "an implementation of the AI RMF functions, categories, and subcategories for a specific setting, application, or technology", and calls itself a cross-sectoral profile because generative models are used across sectors. Two scoping facts matter before anyone uses it as a checklist:

  • It covers only some subcategories. The Profile says "Not every subcategory of the AI RMF is included in this document"; its actions sit under 49 of them. Everything else still comes from the base AI RMF and its Playbook.
  • It covers risks with evidence behind them. NIST says the document "focuses on risks for which there is an existing empirical evidence base", and its working group concentrated on four themes: governance, content provenance, pre-deployment testing and incident disclosure.

What are the 12 GAI risk categories?

The Profile's §2 names 12 risks "unique to or exacerbated by" generative AI. The middle column paraphrases NIST's own definitions; the right column is how each reads for a connected device that calls or embeds a model.

# Risk (NIST name) NIST definition, short Device-maker reading
1 CBRN Information or Capabilities Eased access to or synthesis of information on chemical, biological, radiological or nuclear weapons Rarely in scope for a device feature; relevant if a general chat model is exposed
2 Confabulation Confidently stated but false content A voice assistant or maintenance helper giving wrong instructions
3 Dangerous, Violent, or Hateful Content Violent or inciting content, self-harm or illegal-act recommendations Output filtering on consumer devices
4 Data Privacy Leakage or de-anonymization of personal or sensitive data Audio, video and location data sent to a cloud model
5 Environmental Impacts High compute use in training or operation Mostly a model-provider concern
6 Harmful Bias or Homogenization Amplified bias, performance gaps between groups or languages Speech or vision features that work worse for some users
7 Human-AI Configuration Over-reliance, automation bias, emotional entanglement Operators trusting AI output on safety-relevant equipment
8 Information Integrity Content that blurs fact and fiction, mis- and disinformation Generated reports or summaries presented as fact
9 Information Security Lowered barriers for offensive cyber use; a larger attack surface, including model weights and training data Prompt injection, credential and model-file exposure in firmware
10 Intellectual Property Replication of copyrighted or licensed content, trade secret exposure Licence terms of bundled models and datasets
11 Obscene, Degrading, and/or Abusive Content Synthetic CSAM and non-consensual intimate imagery Image-generation features
12 Value Chain and Component Integration Untraceable third-party components and data, weak supplier vetting Which model, runtime and libraries ship in the image

Twelve is the final count, and the April 2024 initial public draft also had 12. Four names changed between draft and final: "CBRN Information" gained "or Capabilities", "Dangerous or Violent Recommendations" became "Dangerous, Violent, or Hateful Content", "Environmental" became "Environmental Impacts", and "Toxicity, Bias, and Homogenization" became "Harmful Bias or Homogenization". Tools built on the draft can carry the old labels.

How many actions does the Profile suggest, and how are they organized?

The final text holds 212 suggested actions. Each sits in a table under one AI RMF subcategory and carries an Action ID, the action text, and GAI Risks tags naming which of the 12 risks it addresses. Each subcategory table also lists AI Actor Tasks (for example "Governance and Oversight" or "AI Deployment"). NIST explains the ID format itself: "GV-1.1-001 corresponds to the first suggested action for Govern 1.1."

Counting the unique action IDs in the published PDF gives this split:

AI RMF function ID prefix Suggested actions
Govern GV 58
Map MP 39
Measure MS 72
Manage MG 43
Total 212

There is no developer / deployer / end-user tag on individual actions. NIST says instead that "suggested actions relevant to GAI developers may not be relevant to GAI deployers" and leaves the selection to each organization. Most device makers deploy a third-party model rather than train one, which removes many training-data actions from scope but leaves the value-chain, security and incident actions in.

Which suggested actions land on a device or firmware supplier?

When an enterprise or government buyer uses the Profile, its procurement and third-party actions turn into questions for the supplier. These are real action IDs from the published text; the right column is the evidence a supplier can prepare.

Action ID What NIST suggests (paraphrased) Supplier evidence
GV-1.6-003 Inventory entries for GAI systems with data provenance, known issues and related information A component list naming the model, version, runtime and source
GV-2.1-001 Roles and procedures for communicating GAI incidents downstream, including via CVE and NVD A disclosure policy and a security contact that cover the AI feature
GV-6.1-009 Update vendor due diligence to cover IP, privacy, security and "solutions that may rely on embedded GAI technologies" Answers to the buyer's AI questionnaire, with documents
GV-6.2-003 Incident response plans for third-party GAI technologies Who responds when the upstream model provider has an incident
MS-2.7-001 Assess threats such as backdoors, compromised dependencies, model theft or exposure of model weights Firmware and dependency scan results, model-file storage design
MS-2.7-007 AI red-teaming for prompt injection, data poisoning, model extraction and abuse A test report scoped to the shipped feature
MG-3.1-003 Re-assess model risk after fine-tuning or retrieval-augmented generation, and for third-party models used in untested ways Change records for model updates over the air
MG-3.2-002 Document how pre-trained models were adapted A short model card for the shipped model
MG-4.3-003 Report GAI incidents as legal and regulatory requirements demand Mapping to the reporting duties that apply to the product

For the component list, a CycloneDX document can carry models as well as software: the CycloneDX ML-BOM capability describes model and dataset components. See our AI-BOM guide for embedded AI devices for the device-side detail.

What should an AI-enabled device maker hand a buyer who uses AI 600-1?

A short, consistent pack answers most Profile-driven questionnaires without promising more than the product does:

  1. A feature statementWhich features use generative AI, where the model runs (on-device or cloud), and which risk categories from the table above apply. Mark the others "not applicable" with a reason.
  2. A component inventoryThe firmware SBOM plus model entries — name, version, provider, licence — as the firmware SBOM guide describes.
  3. Known-vulnerability statusof the runtime and libraries around the model, with the date of the last check.
  4. Incident and disclosure terms: how a report about the AI feature reaches you, and how you tell customers.
  5. A model-change policy: what triggers re-testing when the model, prompt template or retrieval data changes.

A buyer in the EU may ask the same questions in EU AI Act terms; the EU AI Act cybersecurity guide covers that overlap.

What does an AI 600-1 review look like for one product?

Worked example — synthetic, labelled for illustration. The product, model and findings are made up.

A maker of a building-automation gateway adds a "describe this alarm" feature. The gateway sends alarm context to a cloud LLM and shows the answer on a local web page.

  1. Scope the roleThe maker deploys a third-party model; it does not train one. Training-data actions drop out; GV-6.1-009, GV-6.2-003, MS-2.7-001, MS-2.7-007 and MG-3.1-003 stay in.
  2. Pick the risksConfabulation (wrong alarm advice), Human-AI Configuration (operators trusting it), Information Security (prompt injection through device names an attacker can set) and Value Chain (the API client library and model provider). CBRN, Obscene Content and Environmental Impacts are marked not applicable with a one-line reason each.
  3. InventoryThe firmware SBOM lists the HTTP client, TLS library and JSON parser used by the feature; the model is recorded with provider and version.
  4. TestA red-team pass injects instructions through a BACnet object name; the answer page renders them. The fix escapes device-supplied strings and adds a fixed system instruction.
  5. DocumentAn "AI-generated, verify before acting" label on the page addresses Human-AI Configuration; the change policy says any model version change re-runs step 4.
  6. Answer the buyerThe questionnaire response cites each action ID with the document that covers it.

Is NIST AI 600-1 mandatory — and what changed after Executive Order 14110 was revoked?

NIST AI 600-1 was never mandatory for anyone outside the U.S. federal government, and that hasn't changed. What has changed is the federal policy scaffolding originally built around it:

  • Executive Order 14110 (the order that directed NIST to write the Profile) was revoked on 20 January 2025 by Executive Order 14148. NIST has not withdrawn NIST AI 600-1 in response — the publication remains published and voluntary.
  • OMB Memorandum M-24-10 (28 March 2024) set minimum practices for safety- and rights-impacting AI and encouraged agencies to put "relevant NIST standards, defined pursuant to Sections 4.1(a)" of EO 14110 into generative AI contracts — the section AI 600-1 was later written under. OMB M-25-21 (3 April 2025) "rescinds and replaces" M-24-10 and uses a "high-impact AI" category instead; a companion memo, M-25-22, covers AI acquisition.

Outside federal agencies, the Profile's influence works the way the base AI RMF's does — voluntary adoption, procurement questionnaires that cite NIST AI RMF alignment, and state laws that use it as a benchmark. One documented gap: the Profile has no risk category that squarely addresses autonomous-agent action, although MS-2.7-001 lists "autonomous agents" among threats to assess. That gap is part of why separate agentic-AI guidance has appeared from other organizations, such as OWASP's agentic AI threats and mitigations; treat it as distinct from NIST.

What are the common mistakes when applying the Profile?

  • Treating 212 actions as a checklist. NIST says applicability depends on the organization and its use of GAI; scoring a supplier on all 212 produces noise.
  • Citing draft risk names. "Toxicity, Bias, and Homogenization" is the April 2024 draft label, not the final one.
  • Ignoring the base AI RMF. The Profile adds to AI RMF 1.0; a product review that skips the base framework misses general AI risk.
  • Listing the model but not the code around it. Value Chain risk includes the runtime, client libraries and update path in the firmware, which a model card does not show.
  • Testing once. MG-3.1-003 expects re-assessment when the model or retrieval data changes, which on a connected device can happen without a firmware release.
  • Skipping application-level risks. The Profile's Information Security actions are broad; the OWASP Top 10 for LLM Applications gives concrete LLM attack classes to test against.
  • Assuming the tool layer is covered elsewhere. If the feature calls tools through MCP, the MCP server security guide applies on top.

Where does SoftWSP fit in an AI 600-1 review?

SoftWSP covers one slice of the Profile: the software components around an AI feature. It extracts an SBOM from a firmware image as CycloneDX 1.6 JSON, matches the components against NVD, CISA KEV, FIRST EPSS and CISA CSAF ICS advisories, and drafts a cited report for an engineer to review. That is evidence for the inventory and dependency parts of GV-1.6-003 and MS-2.7-001. It does not test models, red-team prompts or judge confabulation, and an engineer still decides which actions apply. See device & firmware security services or open an SBOM in the SBOM viewer.

This guide summarizes a public NIST publication from its own published text (DOI: 10.6028/NIST.AI.600-1); the risk definitions and action wording in the primary document govern any compliance use. It does not constitute legal or compliance advice.

Put this guide to work

Frequently asked questions

What is NIST AI 600-1, and why was it written?

NIST AI 600-1, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile," is a companion document NIST published on 26 July 2024, developed pursuant to Section 4.1(a)(i)(A) of Executive Order 14110 on Safe, Secure, and Trustworthy Artificial Intelligence, which directed the Secretary of Commerce, acting through NIST, to produce a resource specifically addressing generative AI risk alongside the existing [AI Risk Management Framework](/guides/nist-ai-risk-management-framework/) (AI RMF 1.0 / NIST AI 100-1, published January 2023). NIST describes a Profile as an implementation of the AI RMF's functions, categories and subcategories "for a specific setting, application, or technology"; because generative AI cuts across virtually every sector, NIST designated this one a cross-sectoral profile rather than a use-case-specific one. Per NIST's own description, "the profile can help organizations identify unique risks posed by generative AI and proposes actions for generative AI risk management that best aligns with their goals and priorities." It does not replace AI RMF 1.0 — an organization applying the Generative AI Profile is still expected to address the base AI RMF's general risks on top of it.

What are the 12 GAI risk categories the Profile names?

Per NIST AI 600-1's own text (quoted via corroborated secondary summaries of the published document): CBRN Information or Capabilities ("easier access to information about chemical, biological, radiological, or nuclear weapons and other dangerous materials"); Confabulation ("confidently stated but false content, often called hallucination"); Dangerous, Violent, or Hateful Content; Data Privacy (leakage, unauthorized use or de-anonymization of personal data); Environmental Impacts (the computational/energy cost of training and running generative models); Harmful Bias or Homogenization; Human-AI Configuration (risks from how people interact with and rely on AI, including over-trust and automation bias); Information Integrity (lower barriers to producing misinformation and disinformation at scale); Information Security (new attack surfaces, including prompt injection, plus a bigger attack surface for the AI system itself); Intellectual Property (easier replication of copyrighted or trademarked content, exposure of trade secrets); Obscene, Degrading, and/or Abusive Content; and Value Chain and Component Integration (poorly vetted or untraceable third-party datasets, pre-trained models and components). Twelve is the final published count — the April 2024 initial public draft also listed 12 risks, but under partly different names (for example "Dangerous or Violent Recommendations" and "Toxicity, Bias, and Homogenization"), which NIST renamed for the final July 2024 release.

How is the Profile structured, and how many suggested actions does it contain?

The Profile's actual content is a large catalogue of suggested actions, each tied to one of the AI RMF's four core functions — Govern, Map, Measure, Manage — the same four functions the base AI RMF defines (see our [AI RMF guide](/guides/nist-ai-risk-management-framework/) for what each function covers). Corroborated secondary sources consistently describe 212 suggested actions in the final published text, each carrying an action ID (for example, in the GV-1.1-00x format tying it to a specific AI RMF subcategory) and tagged with the GAI risks it addresses, while each subcategory table lists the pertinent AI Actor Tasks (for example "Governance and Oversight"). The practical effect is less a single checklist than a large menu: an organization is meant to select the actions relevant to its own role in the generative AI value chain (building a model, deploying a third-party model, or using a generative AI product) and its own risk tolerance, not implement all 212 regardless of fit — NIST frames the base AI RMF's actions the same non-mandatory way, and the Profile inherits that framing.

Is NIST AI 600-1 mandatory, and does the 2025 revocation of Executive Order 14110 change anything?

No, it was never mandatory for anyone outside the U.S. federal government, and the revocation of Executive Order 14110 — the order that originally directed NIST to write it, rescinded on 20 January 2025 via the incoming administration's day-one "Initial Rescissions of Harmful Executive Orders and Actions," with Executive Order 14179 following on 23 January 2025 — does not by itself withdraw or invalidate the publication; NIST has not pulled it, and it remains published and voluntary. What has changed is the federal policy layer built on top of it: OMB Memorandum M-24-10 (finalized 28 March 2024) had directed federal agencies to apply minimum risk-management practices to safety- and rights-impacting AI and encouraged them to write NIST standards developed under Section 4.1(a) of Executive Order 14110 — the section the Generative AI Profile was later published under — into generative AI contracts; M-24-10 has since been rescinded and replaced by newer OMB guidance (M-25-21 and a companion M-25-22). Outside the federal government, the Profile's practical weight comes from the same source as the base AI RMF's: voluntary adoption, procurement RFPs, and state AI laws that reference NIST AI RMF alignment as a benchmark, not a legal mandate tied to the Profile itself. One documented gap worth noting: independent analysis of the Profile has pointed out it has no risk category explicitly addressing autonomous-agent action — the risk that an AI agent takes an irreversible action on incorrect reasoning, or that one agent's action cascades into a connected system — which is part of why separate "agentic AI" risk profiles have begun appearing from other organizations rather than from a NIST update to AI 600-1 itself.

Sources

  1. NIST: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1), DOI record
  2. NIST: AI Risk Management Framework (overview page, links to the Generative AI Profile)
  3. NIST AI Resource Center (AIRC): NIST.AI.600-1.GenAI-Profile — initial public draft document
  4. The White House / OMB: Memorandum M-24-10, Advancing Governance, Innovation, and Risk Management for Agency Use of Artificial Intelligence (28 March 2024)
  5. Federal Register: Executive Order 14110, Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence (30 October 2023)
  6. NIST: NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (July 2024), full text PDF — 12 risks (§2), suggested-action tables and table format (§3)
  7. NIST: AI Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (January 2023), full text PDF
  8. Federal Register: Executive Order 14148, Initial Rescissions of Harmful Executive Orders and Actions (signed 20 January 2025)
  9. The White House / OMB: Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust (3 April 2025)
  10. NIST: Dioptra test platform documentation (named in AI 600-1 action MS-2.3-004)
  11. CycloneDX: Machine Learning Bill of Materials (ML-BOM) capability page
Was this guide helpful?

In this cluster

NIST AI 600-1: the Generative AI Profile, explained