Nist ai rmf

<!-- argument-hint: [topic, function, or chapter number] -->

NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

Source: NIST (US Government work, public domain) | Chapters: 8

When to use

Reach for this pack when designing, evaluating, or governing an AI system and needing a structured, standards-grounded approach to risk. Use it to select the right GOVERN/MAP/MEASURE/MANAGE activities for a given lifecycle stage, to understand trustworthiness tradeoffs, or to frame a conversation with leadership about AI risk posture. It is also useful when mapping organisational roles and responsibilities across an AI programme or when setting up a risk profiling exercise.

Prerequisites: none — plain Markdown; no MCP server, API key, or licence tier needed at runtime.

How to Use This Skill

• Without arguments — load the core frameworks below.
• With a topic — ask about trustworthiness characteristics, bias categories, TEVV, third-party risk, AI actor roles, or risk profiles; I read the relevant chapter.
• With a chapter — ask for ch04 (GOVERN), ch05 (MAP), ch06 (MEASURE), ch07 (MANAGE), etc.

Supporting files: glossary.md, patterns.md, cheatsheet.md.

Core Frameworks & Mental Models

The Four Functions (AI RMF Core)

The AI RMF Core organises AI risk management into four functions. GOVERN is the cross-cutting foundation; MAP, MEASURE, and MANAGE operate on specific systems in sequence and iteratively.

GOVERN (cross-cutting, always on)
  └── MAP (context & risk identification)
        └── MEASURE (quantify & evaluate)
              └── MANAGE (respond, monitor, improve)

GOVERN establishes organisational culture, policies, accountability structures, workforce requirements, and third-party governance. Its six categories cover: (1) policies and legal/regulatory compliance, (2) accountability and training, (3) team diversity and human-AI configuration, (4) safety-first risk culture, (5) external stakeholder feedback, (6) third-party and supply chain risk. GOVERN must be in place before MAP activities are meaningful.

MAP establishes the contextual knowledge needed to frame risks: intended purposes, deployment setting, legal norms, stakeholder population, system capabilities, cost-benefit analysis, third-party components, and societal impacts. MAP 1–5 categories build from context (MAP 1) through impact characterisation (MAP 5). MAP outputs enable the go/no-go decision and are direct inputs to MEASURE.

MEASURE applies quantitative, qualitative, or mixed-method tools to evaluate trustworthiness characteristics. Its thirteen MEASURE 2.x subcategories map one-to-one to the trustworthiness characteristics plus environmental impact and TEVV effectiveness. Key structural requirement: assessors must be independent from builders (MEASURE 1.3). All measurement must be documented, including risks that cannot currently be measured.

MANAGE allocates resources to prioritised risks, executes risk treatment plans (mitigate / transfer / avoid / accept — MANAGE 1.3), documents residual risks for downstream parties (MANAGE 1.4), monitors deployed systems (MANAGE 4.1), and communicates incidents to affected communities (MANAGE 4.3). A deactivation mechanism (MANAGE 2.4) must be in place before deployment.

Seven Trustworthiness Characteristics

Valid and Reliable is the necessary base; without it, no other characteristic can be meaningfully claimed. Accountable and Transparent is cross-cutting and applies to all others.

1. Valid & Reliable — system performs as required, confirmed by objective evidence, across its intended lifecycle.
2. Safe — no conditions under which life, health, property, or environment is endangered.
3. Secure & Resilient — withstands attacks and adverse events; degrades gracefully.
4. Accountable & Transparent — information available to those interacting with the system; enables redress.
5. Explainable & Interpretable — mechanisms of operation described ("how"); output meaning communicated ("why").
6. Privacy-Enhanced — autonomy, identity, and dignity protected through design choices.
7. Fair / Bias Managed — equality and equity addressed; three bias categories managed (systemic, computational/statistical, human-cognitive).

Tradeoffs among characteristics are normal and require documented, contextual human judgement — the framework does not resolve them algorithmically.

Risk Framing

• Risk = probability × magnitude of harm (ISO 31000:2018 / OMB A-130:2016)
• Risk tolerance is contextual and organisation-specific; the AI RMF does not prescribe it.
• Residual risk must be documented and disclosed to downstream acquirers and end users.
• When risk is unacceptable (severe harm imminent), halt development or deployment until it can be managed.

AI RMF Profiles

Current Profile (existing state) → Gap Analysis → Target Profile (desired state) → Action Plan. This is the mechanism for turning the generic framework into organisation-specific improvement programmes. Cross-sectoral profiles address capabilities (e.g., LLMs) used across multiple sectors.

Chapter Index

#	Section	Key content
ch01	Framing AI Risk	Risk definition, risk tolerance, risk prioritisation, socio-technical risk, measurement challenges
ch02	Audience and AI Actors	AI actors, OECD lifecycle dimensions, TEVV, diverse teams, affected communities
ch03	Trustworthiness Characteristics	All seven characteristics, three bias categories, tradeoff analysis, measurement obligations
ch04	GOVERN Function	GOVERN 1–6, risk culture, accountability, diversity, third-party governance, decommissioning
ch05	MAP Function	MAP 1–5, context establishment, system categorisation, impact characterisation, go/no-go
ch06	MEASURE Function	MEASURE 1–4, TEVV, independent assessment, MEASURE 2.1–2.13 subcategory details
ch07	MANAGE Function	MANAGE 1–4, risk response options, residual risk, deactivation, incident communication
ch08	Profiles and Use	Current/Target profiles, gap analysis, cross-sectoral profiles, AI RMF Playbook, living document

Topic Index

• Accountability → ch03, ch04 (GOVERN 2)
• Bias (three categories) → ch03, ch06 (MEASURE 2.11)
• Decommissioning → ch04 (GOVERN 1.7), ch07 (MANAGE 4.1)
• Deactivation mechanism → ch07 (MANAGE 2.4)
• Diverse teams → ch02, ch04 (GOVERN 3), ch05 (MAP 1.2)
• Environmental impact → ch06 (MEASURE 2.12)
• Explainability / Interpretability → ch03, ch06 (MEASURE 2.9)
• Fairness → ch03, ch06 (MEASURE 2.11)
• Go/no-go decision → ch05 (MAP), ch07 (MANAGE 1.1)
• GOVERN categories → ch04
• Human-AI interaction → ch02, ch04 (GOVERN 3.2), ch05 (MAP 3.5)
• Incident response → ch07 (MANAGE 4.3)
• MAP categories → ch05
• MANAGE categories → ch07
• MEASURE subcategories → ch06
• Privacy → ch03, ch06 (MEASURE 2.10)
• Profiles (Current/Target) → ch08
• Residual risk → ch01, ch07 (MANAGE 1.4)
• Risk measurement challenges → ch01
• Risk tolerance → ch01, ch04 (GOVERN 1.3), ch05 (MAP 1.5)
• Robustness / generalizability → ch03
• Safety → ch03, ch06 (MEASURE 2.6)
• Security and resilience → ch03, ch06 (MEASURE 2.7)
• Socio-technical risk → ch01, ch02
• TEVV → ch02, ch06 (MEASURE 1.3, MEASURE 2.1–2.2)
• Third-party risk → ch01, ch04 (GOVERN 6), ch05 (MAP 4), ch07 (MANAGE 3)
• Transparency → ch03, ch06 (MEASURE 2.8)
• Trustworthiness characteristics → ch03
• Valid and Reliable → ch03, ch06 (MEASURE 2.5)

Supporting Files

• glossary.md — ~45 terms with chapter references, alphabetical.
• patterns.md — 13 actionable patterns with when-to-use and trade-offs.
• cheatsheet.md — decision rules, function quick-reference table, trustworthiness characteristic table, tells and smells.

---

Scope & Limits

This pack covers the NIST AI RMF 1.0 (NIST AI 100-1, January 2023) in full: foundational framing, audience, trustworthiness characteristics, all four Core functions and their categories/subcategories, and profiles. It is a US Government publication and is public domain. This pack does not cover: the AI RMF Playbook's tactical subcategory suggestions (separate online resource), the EU AI Act, ISO/IEC 42001 in depth, sector-specific regulations, or post-1.0 NIST AI RMF updates. For security and privacy implementation detail, consult the NIST Cybersecurity Framework and NIST Privacy Framework packs directly.