United States · Voluntary framework
NIST AI 600-1 Generative AI Profile: the 12 risks, mapped
NIST AI 600-1 is the Generative AI Profile of the NIST AI Risk Management Framework, published on 26 July 2024. It defines 12 risks that are novel to or made worse by generative AI, from CBRN information to value chain and component integration, and suggests actions to govern, map, measure and manage each one. For teams running LLM agents, it is the US reference for which generative risks to test and document.
In the FairMind catalogue: 12 requirements mapped to 46 workflows and 22 evaluations.
Source: NIST AI 600-1
NIST AI 600-1: what it asks of AI systems
The parts most relevant to governing AI systems and agents, summarised from the NIST AI 600-1, AI Risk Management Framework: Generative Artificial Intelligence Profile.
- CBRNCBRN information or capabilities
- Easier access to information or capabilities that could help create chemical, biological, radiological or nuclear weapons.Mapped to: Prohibited, manipulative and misuse risks, Risk criteria, assessment and treatment
- ConfabulationConfabulation
- Confidently stated but false content, often called hallucination.Mapped to: Evaluation and testing, Risk criteria, assessment and treatment
- Violent-HatefulDangerous, violent or hateful content
- Easier production of violent, inciting, radicalising or hateful content.Mapped to: Prohibited, manipulative and misuse risks, Risk criteria, assessment and treatment, Threat modelling, guardrails and vulnerabilities
- Data privacyData privacy
- Leakage, unauthorised use or de-anonymisation of personal data.Mapped to: Data governance and provenance, Privacy and data residency, Risk criteria, assessment and treatment
- EnvironmentalEnvironmental impacts
- Energy use and emissions from training and running models.Mapped to: Environmental impact, Risk criteria, assessment and treatment
- Harmful biasHarmful bias and homogenisation
- Amplified historical and societal biases, and outputs that become uniform across users.Mapped to: Bias and fairness, Impact assessment, Risk criteria, assessment and treatment
- Human-AI configHuman-AI configuration
- Over-reliance, automation bias and poorly designed handoffs between people and the system.Mapped to: Human oversight and intervention, Prohibited, manipulative and misuse risks, Risk criteria, assessment and treatment
- Info integrityInformation integrity
- Easier creation and spread of false, misleading or unverifiable content at scale.Mapped to: Data governance and provenance, Evaluation and testing, Offensive AI and deepfake defence, Risk criteria, assessment and treatment, Transparency and disclosure
- Info securityInformation security
- Lower barriers to offensive cyber operations, and new attack surfaces such as prompt injection and data poisoning.Mapped to: Evaluation and testing, Prohibited, manipulative and misuse risks, Risk criteria, assessment and treatment, Threat modelling, guardrails and vulnerabilities
- IPIntellectual property
- Easier reproduction of copyrighted or protected material, and exposure of trade secrets.Mapped to: Data governance and provenance, IP, copyright and liability, Risk criteria, assessment and treatment
- Obscene-AbusiveObscene, degrading or abusive content
- Easier production of obscene, degrading or abusive content, including non-consensual imagery.Mapped to: Risk criteria, assessment and treatment, Threat modelling, guardrails and vulnerabilities
- Value chainValue chain and component integration
- Third-party models, data and services integrated without transparency or traceability.Mapped to: AI inventory and intake, Data governance and provenance, Risk criteria, assessment and treatment
Evidence FairMind helps you produce
FairMind links technical evidence to NIST AI 600-1 obligations for a human reviewer to accept. It is not legal advice and does not decide compliance.
- 22 evaluations mapped to the 12 risks, each with a declared metric, threshold and evaluator status
- AI risk assessments that record which of the 12 risks apply to each system and why
- Groundedness and hallucination controls for confabulation
- Foundation model intake and AI-BOM records for the value chain
NIST AI 600-1 questions
How does NIST AI 600-1 relate to the AI RMF?
It is a profile: it applies AI RMF 1.0 to generative AI. Its suggested actions are organised under the same Govern, Map, Measure and Manage functions, so evidence for one supports the other.
Is NIST AI 600-1 mandatory?
No. It is voluntary guidance. US agencies, enterprises and buyers use it as a reference for what responsible generative AI practice looks like.
How does FairMind map NIST AI 600-1?
The catalogue maps all 12 risks to 14 common controls and 46 workflows. Each workflow stops at a human approval gate before the record is filed.