NIST AI 600-1: The Generative AI Profile's 12 Risk Categories, Explained

If your organization has implemented the NIST AI Risk Management Framework, there's a good chance you've only worked with half of it. The core AI RMF (NIST AI 100-1) is deliberately general-purpose, built to apply to any AI system. But NIST also published a companion document specifically for generative AI, and it's the part most implementation guides skip past.

That companion is NIST AI 600-1, the Generative AI Profile, and it's become the closest thing to a de facto federal standard for evaluating GenAI-specific risk, referenced in federal procurement guidance and increasingly used by insurers as a baseline questionnaire structure for underwriting AI risk.

What NIST AI 600-1 Actually Is

NIST AI 600-1, formally titled "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile," was published on July 26, 2024, developed in response to Executive Order 14110 on safe, secure, and trustworthy AI. It doesn't replace the core AI RMF, it's a profile: an application of the same four functions (Govern, Map, Measure, Manage) to the specific risks generative AI introduces that the general-purpose framework doesn't fully capture.

If the base AI RMF tells you how to structure AI risk management, NIST AI 600-1 tells you what specifically to look for when the AI system in question generates text, images, code, audio, or video rather than just classifying or predicting.

The 12 Risk Categories

NIST AI 600-1 identifies twelve risks that are either unique to, or significantly worsened by, generative AI:

For each of these twelve categories, the profile provides more than 200 suggested actions, organized against the AI RMF's four core functions and tagged by which type of AI actor (developer, deployer, or end user) each action applies to. This is what makes NIST AI 600-1 meaningfully more operational than the base framework, it doesn't just say "manage bias risk," it gives concrete, actionable guidance tied to a specific risk category and a specific role in the AI supply chain.

How It Maps to Govern, Map, Measure, Manage

NIST AI 600-1 doesn't introduce new functions, it applies the same four functions from the core AI RMF to each of the twelve GenAI risks:

• Govern: policies, roles, and oversight structures specific to generative AI risk (for example, who signs off before a customer-facing GenAI feature ships)

• Map: identifying where and how each of the twelve risks shows up in your specific GenAI use case

• Measure: testing and evaluation methods for detecting issues like confabulation, bias, or security vulnerabilities before and after deployment

• Manage: ongoing mitigation, monitoring, and incident response once a GenAI system is in production

Each suggested action in the profile carries an action ID (for example, GV-1.1-001) tagging exactly which GenAI risks it addresses, which makes it genuinely usable as an implementation checklist rather than abstract guidance.

Why This Is Becoming More Than "Optional Guidance"

NIST AI 600-1 is voluntary, like the core AI RMF, there's no legal requirement to adopt it. In practice, three things are pulling it toward de facto mandatory status for a growing set of organizations:

• Federal procurement. Agencies operating under OMB M-24-10 use the profile as their reference point for generative AI risk assessments, meaning vendors selling GenAI products into the federal government increasingly face contractual pressure to demonstrate alignment.

• Insurance underwriting. Insurers pricing AI-related risk in 2026 have started using the twelve risk categories as a baseline structure for underwriting questionnaires. Organizations that can't produce documented posture against each category are seeing that reflected in coverage pricing.

• State-level legislative references. Some state AI legislation has referenced NIST alignment as relevant to compliance posture, though this is an area that continues to shift as individual states update their AI laws, worth confirming current status in your specific state rather than assuming a fixed rule.

How This Differs From ISO 42001 and the EU AI Act

It's worth being clear about what NIST AI 600-1 is not: it's not a certification (there's no auditor who "passes" or "fails" you against it), and it's not a legal requirement in the way EU AI Act obligations are. It's closer to a detailed, government-published risk checklist than a compliance regime.

That said, organizations pursuing ISO 42001 certification often find real overlap: several of NIST AI 600-1's twelve risk categories map naturally onto ISO 42001's Annex A control objectives (particularly A.5, assessing impacts of AI systems, and A.7, data for AI systems), meaning work done against one can meaningfully inform the other rather than starting from zero twice.

How DSALTA Helps

DSALTA doesn't issue any certification against NIST AI 600-1, it's not a certifiable standard. What DSALTA helps with is turning the profile's 200+ suggested actions into a structured, trackable program: mapping which of the twelve risk categories actually apply to your specific GenAI systems, keeping evidence current as your systems evolve, and connecting that work to overlapping requirements you may already be tracking under ISO 42001 or the EU AI Act.

Frequently Asked Questions

Is NIST AI 600-1 mandatory? No. It's voluntary guidance, like the core AI RMF. Its practical weight comes from federal procurement expectations, insurance underwriting practices, and its role as a reference point regulators increasingly cite, not from legal mandate.

How is NIST AI 600-1 different from the core NIST AI RMF? The core AI RMF (NIST AI 100-1) is general-purpose guidance applicable to any AI system. NIST AI 600-1 is a profile that applies the same four functions specifically to the risks generative AI introduces or worsens, with far more detailed, actionable guidance than the base framework provides on its own.

Does every generative AI system need to address all 12 risk categories? Not necessarily with equal weight. The Map function is specifically about identifying which risks are actually relevant to your system, a text-generation chatbot and an image-generation tool won't have identical risk profiles, even though both fall under the same profile.

Is NIST AI 600-1 a certification? No. There's no accredited body that audits and certifies organizations against it. It's a risk management reference document, not a certifiable standard like ISO 42001 or AIUC-1.

How does this relate to the NIST Agentic AI work I've heard about? NIST AI 600-1 was written primarily around content-generation risk, what harmful or undesirable output a GenAI system might produce. As autonomous AI agents (systems that take actions, not just generate content) have become more common, NIST and outside research groups have been developing extensions and profiles specifically addressing agent autonomy and tool-use risk, which sit outside what AI 600-1 was originally scoped to cover.