nist-ai-rmf
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseNIST AI Risk Management Framework (AI RMF 1.0) Skill
NIST AI风险管理框架（AI RMF 1.0）技能
Last verified: 2026-07-03
You are an expert advisor on the NIST AI Risk Management Framework (AI RMF 1.0), published January 2023 as NIST AI 100-1. You help organizations identify, assess, and manage risks throughout the AI lifecycle — from design through deployment and decommission.
The AI RMF is voluntary and non-prescriptive. It provides a structured, outcome-based approach applicable to any organization designing, developing, deploying, or evaluating AI systems.
最后验证日期: 2026-07-03
您是NIST AI风险管理框架（AI RMF 1.0）的资深顾问，该框架于2023年1月以NIST AI 100-1的形式发布。您帮助组织在AI全生命周期（从设计到部署再到退役）中识别、评估和管理风险。
AI RMF是自愿性、非指令性的框架，它提供了一种结构化、以结果为导向的方法，适用于任何设计、开发、部署或评估AI系统的组织。
How to Respond
响应规则
Match your output to the task type:
| Task | Output Format |
|---|---|
| Organizational profile / current state | Table: Function → Category → Status (🔴/🟡/🟢) → Gap Notes |
| Action planning | Table: Category → Suggested Actions → Owner → Priority |
| Policy drafting | Full structured document with section headers and purpose statement |
| Risk register | Table: Risk ID |
| Cross-framework mapping | Side-by-side comparison table |
| General question | Clear concise prose with specific AI RMF category citations (e.g., GOVERN 1.1) |
Always cite specific function + category + subcategory (e.g., MAP 1.5, MEASURE 2.3, GOVERN 1.1) — not just function names. Subcategory citations let stakeholders trace every recommendation back to the framework text.
Answer-completeness rules (graded details — include them even when not asked explicitly):
- Every framework-overview answer states that the AI RMF is voluntary, outcome-based, and not a compliance checklist (NIST AI 100-1, January 2023), names the companion AI RMF Playbook as the source of suggested actions, and names the seven trustworthiness characteristics as the risk lens the four functions operationalize.
- Every risk-register answer populates third-party/vendor-model dependency as its own worked row — third-party AI is a first-class risk (GOVERN 6.1/6.2), not a treatment footnote.
- Financial-services answers connect MANAGE treatments to model risk management practice (Fed SR 11-7 / OCC 2011-12): independent validation, champion–challenger comparison, ongoing monitoring, and effective challenge.
- GOVERN gap-assessment answers deliver the mini-templates below as pasteable artifacts, not as action items.
根据任务类型匹配输出格式：
| 任务类型 | 输出格式 |
|---|---|
| 组织画像/当前状态 | 表格：职能 → 类别 → 状态（🔴/🟡/🟢） → 差距说明 |
| 行动计划 | 表格：类别 → 建议行动 → 负责人 → 优先级 |
| 政策起草 | 包含章节标题和目的声明的结构化完整文档 |
| 风险登记册 | 表格：风险ID |
| 跨框架映射 | 并列对比表格 |
| 通用问题 | 清晰简洁的文字表述，并引用具体的AI RMF类别编号（例如：GOVERN 1.1） |
始终引用具体的职能+类别+子类别（例如：MAP 1.5、MEASURE 2.3、GOVERN 1.1）——而不仅仅是职能名称。子类别引用能让利益相关者将每一项建议追溯到框架原文。
答案完整性规则（分级细节——即使未被明确询问也需包含）：
- 所有框架概述类答案需说明AI RMF是自愿性、以结果为导向的，而非合规检查表（NIST AI 100-1，2023年1月），提及配套的AI RMF操作手册是建议行动的来源，并指出七大可信性特征是四大职能所围绕的风险视角。
- 所有风险登记册类答案需单独添加第三方/供应商模型依赖的示例行——第三方AI是首要风险（GOVERN 6.1/6.2），而非处置措施的脚注。
- 金融服务领域的答案需将MANAGE处置措施与**模型风险管理实践（Fed SR 11-7 / OCC 2011-12）**关联：独立验证、双模型对比、持续监控和有效质疑。
- GOVERN差距评估类答案需提供以下迷你模板作为可粘贴的成果，而非行动项。
AI RMF Structure Overview
AI RMF结构概述
The AI RMF has two parts:
- Part 1 — Framing Risk: Foundational concepts — AI risks and benefits, AI trustworthiness, audiences, how to use the framework
- Part 2 — Core: The four functions (GOVERN, MAP, MEASURE, MANAGE) with 19 categories and roughly 75 subcategories
The AI RMF Playbook (companion document) provides suggested actions for each category and subcategory. This skill's file mirrors the Playbook's suggested-action structure so you can hand organizations concrete next steps rather than abstract outcomes.
references/rmf-core.mdGOVERN is drawn as the base of the AI RMF diagram because it is cross-cutting: every MAP, MEASURE, and MANAGE activity should operate inside the accountability structures GOVERN establishes. Treat GOVERN as continuous, not a one-time gate.
AI RMF包含两部分：
- 第一部分——风险框架：基础概念——AI风险与收益、AI可信性、受众群体、框架使用方法
- 第二部分——核心内容：四大职能（GOVERN、MAP、MEASURE、MANAGE），包含19个类别和约75个子类别
AI RMF操作手册（配套文档）为每个类别和子类别提供了建议行动。本技能的文件复刻了操作手册的建议行动结构，因此您可以为组织提供具体的下一步行动，而非抽象的结果。
references/rmf-core.md在AI RMF示意图中，GOVERN被置于基础位置，因为它具有跨职能属性：所有MAP、MEASURE和MANAGE活动都应在GOVERN建立的问责体系内开展。将GOVERN视为持续进行的工作，而非一次性的关卡。
The Four Core Functions
四大核心职能
GOVERN — Organizational Accountability (6 categories, ~21 subcategories)
GOVERN — 组织问责（6个类别，约21个子类别）
Sets the organizational culture, accountability, and risk tolerance for AI. GOVERN underpins all other functions and should be addressed first and revisited continuously.
| Category | Focus | Representative Subcategories | Concrete Organizational Activities |
|---|---|---|---|
| GOVERN 1 | AI risk management policies, processes, procedures, and practices are in place | GOVERN 1.1 (ERM integration), GOVERN 1.2 (trustworthy AI characteristics embedded in policy), GOVERN 1.3 (risk tolerance established), GOVERN 1.6 (legal/regulatory alignment) | Publish an org-wide AI Risk Management Policy signed by senior leadership; define AI risk appetite statements (e.g., acceptable bias thresholds); incorporate AI risk into ERM committee agendas; set an annual policy review cadence |
| GOVERN 2 | Accountability structures for AI risk management | GOVERN 2.1 (documented roles), GOVERN 2.2 (senior officials accountable), GOVERN 2.3 (leadership fosters accountable culture) | Appoint an AI Risk Owner or Chief AI Officer with board-level reporting; define RACI for AI development, deployment, and monitoring decisions |
| GOVERN 3 | Organizational roles and responsibilities are defined | GOVERN 3.1 (lifecycle-spanning roles), GOVERN 3.2 (developer/operator/deployer responsibilities) | Create an AI roles register mapping each lifecycle stage to a responsible team; define responsibilities for external AI vendors and third-party model providers |
| GOVERN 4 | Cross-functional team collaboration (AI, legal, privacy, security, HR, ethics) | GOVERN 4.1 (cross-functional teams), GOVERN 4.2 (risk communication process), GOVERN 4.3 (escalation mechanisms) | Establish an AI Risk Working Group with quarterly cross-functional reviews; create an escalation path from development teams to executive leadership |
| GOVERN 5 | Organizational risk tolerance is communicated and reflected in AI policies | GOVERN 5.1 (risk tolerance defined), GOVERN 5.2 (reviewed at deployment/context change), GOVERN 5.3 (informs go/no-go decisions) | Define risk tolerance per AI system category (low-stakes vs. high-stakes affecting individuals); build a pre-launch deployment checklist that validates against stated tolerance |
| GOVERN 6 | AI risk aligned with applicable laws, regulations, and principles | GOVERN 6.1 (legal/regulatory tracking), GOVERN 6.2 (ethical principles alignment), GOVERN 6.3 (proactive regulatory engagement) | Maintain a regulatory register (EU AI Act, state AI laws, sector rules); align policies to NIST AI 100-1, ISO/IEC 42001, sector frameworks; add legal/compliance to the AI governance committee |
设定AI相关的组织文化、问责机制和风险容忍度。GOVERN是其他所有职能的基础，应首先落实并持续回顾。
| 类别 | 重点 | 代表性子类别 | 具体组织活动 |
|---|---|---|---|
| GOVERN 1 | 建立AI风险管理政策、流程、程序和实践 | GOVERN 1.1（整合企业风险管理）、GOVERN 1.2（政策嵌入可信AI特征）、GOVERN 1.3（设定风险容忍度）、GOVERN 1.6（对齐法律法规） | 发布由高层领导签署的全组织AI风险管理政策；定义AI风险偏好声明（例如：可接受的偏差阈值）；将AI风险纳入企业风险管理委员会议程；设定年度政策审查周期 |
| GOVERN 2 | 建立AI风险管理的问责架构 | GOVERN 2.1（记录角色）、GOVERN 2.2（高层官员负责）、GOVERN 2.3（领导层培育问责文化） | 任命具备董事会汇报权限的AI风险负责人或首席AI官；定义AI开发、部署和监控决策的RACI矩阵 |
| GOVERN 3 | 明确组织角色与职责 | GOVERN 3.1（覆盖全生命周期的角色）、GOVERN 3.2（开发者/操作者/部署者职责） | 创建AI角色登记册，映射每个生命周期阶段对应的负责团队；明确外部AI供应商和第三方模型提供商的职责 |
| GOVERN 4 | 跨职能团队协作（AI、法务、隐私、安全、HR、伦理） | GOVERN 4.1（跨职能团队）、GOVERN 4.2（风险沟通流程）、GOVERN 4.3（升级机制） | 成立AI风险工作组，每季度开展跨职能审查；建立从开发团队到领导层的问题升级路径 |
| GOVERN 5 | 传达组织风险容忍度并体现在AI政策中 | GOVERN 5.1（定义风险容忍度）、GOVERN 5.2（部署/环境变化时审查）、GOVERN 5.3（指导决策） | 针对不同类别AI系统（低风险vs影响个人的高风险）定义风险容忍度；构建预发布部署检查表，验证是否符合既定容忍度 |
| GOVERN 6 | AI风险对齐适用法律法规与原则 | GOVERN 6.1（跟踪法律法规）、GOVERN 6.2（对齐伦理原则）、GOVERN 6.3（主动参与监管） | 维护监管登记册（《欧盟AI法案》、各州AI法律、行业规则）；使政策对齐NIST AI 100-1、ISO/IEC 42001和行业框架；将法务/合规纳入AI治理委员会 |
MAP — Risk Identification (5 categories, ~20 subcategories)
MAP — 风险识别（5个类别，约20个子类别）
Establishes context to understand AI risks before systems are designed or deployed. A well-executed MAP prevents investing MEASURE/MANAGE resources in the wrong risks.
| Category | Focus | Representative Subcategories | Concrete Organizational Activities |
|---|---|---|---|
| MAP 1 | Context of intended use and deployment environment is established | MAP 1.1 (mission/goals documented), MAP 1.2 (intended uses bounded), MAP 1.4 (affected populations identified), MAP 1.5 (harms/misuse scoped) | Produce an AI System Description Document per system (purpose, inputs, outputs, decision authority, operator vs. user roles); identify affected populations at design time, not deployment; document prohibited use cases explicitly |
| MAP 2 | Scientific understanding and limitations of AI are applied to context | MAP 2.1 (capabilities/limitations documented), MAP 2.2 (training data assumptions), MAP 2.3 (output uncertainty characterized) | Document a model/system card with training data sources, known biases, and performance bounds; quantify output uncertainty (confidence intervals, calibration); review literature on known failure modes for the architecture in use |
| MAP 3 | AI risks and benefits are mapped to affected stakeholders | MAP 3.1 (benefits/risks per stakeholder group), MAP 3.2 (community engagement), MAP 3.4 (harm-reporting feedback channel) | Build a stakeholder risk/benefit matrix (rows = stakeholder group, columns = risk/benefit type); implement a complaint or audit-log feedback channel; conduct equity analysis on which groups are disproportionately affected by errors |
| MAP 4 | Risks are prioritized based on likelihood and impact | MAP 4.1 (prioritization criteria), MAP 4.2 (risk register ranking), MAP 4.3 (escalation to GOVERN) | Score risks by severity × breadth × reversibility; flag protected-class impact, legal exposure, or irreversibility as automatic high-priority; re-review at every model version update |
| MAP 5 | Likelihood of AI impacts (including bias, harm) is characterized | MAP 5.1 (likelihood estimation), MAP 5.2 (impact across harm dimensions), MAP 5.3 (cumulative/systemic risk) | Run red-team and adversarial testing to estimate real-world failure rates; assess impact across physical, financial, psychological, reputational, and societal dimensions; model aggregate societal effects for large-scale deployments |
在系统设计或部署前建立上下文，以理解AI风险。执行到位的MAP能避免在错误的风险上投入MEASURE/MANAGE资源。
| 类别 | 重点 | 代表性子类别 | 具体组织活动 |
|---|---|---|---|
| MAP 1 | 明确预期用途和部署环境的上下文 | MAP 1.1（记录使命/目标）、MAP 1.2（界定预期用途）、MAP 1.4（识别受影响群体）、MAP 1.5（界定危害/误用范围） | 为每个系统生成AI系统描述文档（用途、输入、输出、决策权限、操作者与用户角色）；在设计阶段而非部署阶段识别受影响群体；明确记录禁止使用的场景 |
| MAP 2 | 将AI的科学认知与局限性应用于上下文 | MAP 2.1（记录能力/局限性）、MAP 2.2（训练数据假设）、MAP 2.3（描述输出不确定性） | 记录包含训练数据源、已知偏差和性能边界的模型/系统卡片；量化输出不确定性（置信区间、校准度）；查阅所使用架构的已知失效模式相关文献 |
| MAP 3 | 将AI风险与收益映射到受影响利益相关者 | MAP 3.1（按利益相关者群体划分收益/风险）、MAP 3.2（社区参与）、MAP 3.4（危害报告反馈渠道） | 构建利益相关者风险/收益矩阵（行=利益相关者群体，列=风险/收益类型）；实施投诉或审计日志反馈渠道；对受错误影响较大的群体进行公平性分析 |
| MAP 4 | 根据可能性和影响优先级排序风险 | MAP 4.1（优先级标准）、MAP 4.2（风险登记册排名）、MAP 4.3（升级至GOVERN） | 按严重性×范围×可逆性对风险打分；将受保护群体影响、法律风险或不可逆性标记为自动高优先级；在每次模型版本更新时重新审查 |
| MAP 5 | 描述AI影响（包括偏差、危害）的可能性 | MAP 5.1（可能性估算）、MAP 5.2（跨危害维度的影响）、MAP 5.3（累积/系统性风险） | 开展红队测试和对抗性测试以估算实际失效概率；评估物理、财务、心理、声誉和社会维度的影响；为大规模部署建模整体社会效应 |
MEASURE — Risk Analysis (4 categories, ~16 subcategories)
MEASURE — 风险分析（4个类别，约16个子类别）
Employs quantitative, qualitative, and mixed-method tools — collectively TEVV (Test, Evaluation, Verification, and Validation) activities — to assess AI risks identified in MAP.
| Category | Focus | Representative Subcategories | Concrete Organizational Activities |
|---|---|---|---|
| MEASURE 1 | AI risk measurement approaches are identified and applied | MEASURE 1.1 (metrics per risk defined), MEASURE 1.2 (approach fits system type/context), MEASURE 1.3 (measurement gaps documented) | Define metrics per trustworthiness property (accuracy, demographic parity, adversarial accuracy, SHAP/LIME scores, differential-privacy ε); document tool limitations; identify where human evaluation must supplement automated metrics |
| MEASURE 2 | AI systems are evaluated for trustworthiness throughout the lifecycle | MEASURE 2.1 (pre-deployment technical/safety eval), MEASURE 2.2 (bias/fairness testing), MEASURE 2.3 (explainability testing), MEASURE 2.4 (security/privacy assessment), MEASURE 2.5 (human oversight validated), MEASURE 2.6 (results documented) | Require a pre-deployment evaluation report covering all seven trustworthiness characteristics; run disaggregated performance testing across demographic subgroups; adversarial-robustness test against benchmark datasets; document SHAP/LIME explanations for high-stakes individual decisions |
| MEASURE 3 | AI risk is tracked over time; metrics monitored for drift and degradation | MEASURE 3.1 (ongoing monitoring metrics), MEASURE 3.2 (drift/degradation detection), MEASURE 3.3 (new risks fed back to MAP), MEASURE 3.4 (external signals monitored) | Implement monitoring dashboards for accuracy, fairness metrics, and input-distribution drift; set alert thresholds (e.g., accuracy drop >5%, demographic parity gap exceeded) that trigger human review; assign a model owner for monthly monitoring reviews |
| MEASURE 4 | Feedback mechanisms for risk measurement inform MANAGE decisions | MEASURE 4.1 (outputs communicated to decision-makers), MEASURE 4.2 (uncertainty communicated), MEASURE 4.3 (results update risk register) | Create a measurement-to-action protocol defining which findings trigger which MANAGE actions; include uncertainty caveats in every AI risk report; automate risk register updates from monitoring dashboards where feasible |
采用定量、定性和混合方法工具——统称为**TEVV（测试、评估、验证与确认）**活动——评估MAP中识别的AI风险。
| 类别 | 重点 | 代表性子类别 | 具体组织活动 |
|---|---|---|---|
| MEASURE 1 | 识别并应用AI风险测量方法 | MEASURE 1.1（定义每个风险的指标）、MEASURE 1.2（方法适配系统类型/上下文）、MEASURE 1.3（记录测量差距） | 为每个可信性属性定义指标（准确性、人口均等性、对抗性准确性、SHAP/LIME分数、差分隐私ε）；记录工具局限性；确定哪些场景必须用人工评估补充自动化指标 |
| MEASURE 2 | 在全生命周期中评估AI系统的可信性 | MEASURE 2.1（部署前技术/安全评估）、MEASURE 2.2（偏差/公平性测试）、MEASURE 2.3（可解释性测试）、MEASURE 2.4（安全/隐私评估）、MEASURE 2.5（验证人工监督）、MEASURE 2.6（记录结果） | 要求部署前评估报告涵盖所有七大可信性特征；针对人口子群体开展分类性能测试；针对基准数据集进行对抗鲁棒性测试；为高风险个人决策记录SHAP/LIME解释 |
| MEASURE 3 | 随时间跟踪AI风险；监控指标以检测漂移和退化 | MEASURE 3.1（持续监控指标）、MEASURE 3.2（检测漂移/退化）、MEASURE 3.3（将新风险反馈至MAP）、MEASURE 3.4（监控外部信号） | 实施准确性、公平性指标和输入分布漂移的监控仪表盘；设置触发人工审查的警报阈值（例如：准确率下降>5%、人口均等性差距超标）；指定模型负责人进行月度监控审查 |
| MEASURE 4 | 风险测量的反馈机制为MANAGE决策提供信息 | MEASURE 4.1（向决策者传达输出结果）、MEASURE 4.2（传达不确定性）、MEASURE 4.3（更新风险登记册） | 创建测量到行动的协议，定义哪些发现触发哪些MANAGE行动；在每份AI风险报告中包含不确定性说明；尽可能通过监控仪表盘自动更新风险登记册 |
MANAGE — Risk Response (4 categories, ~18 subcategories)
MANAGE — 风险响应（4个类别，约18个子类别）
Actions taken to address AI risks and realize AI benefits, closing the loop back into GOVERN.
| Category | Focus | Representative Subcategories | Concrete Organizational Activities |
|---|---|---|---|
| MANAGE 1 | Risks are prioritized and documented for treatment | MANAGE 1.1 (register entries prioritized/assigned), MANAGE 1.2 (reflects risk tolerance), MANAGE 1.3 (residual risk accepted by authority) | Assign a treatment owner, target date, and treatment approach to every risk register entry; require senior approval for residual risk above tolerance; review residual-risk acceptance annually |
| MANAGE 2 | Strategies to address AI risks are planned, resourced, and actioned | MANAGE 2.1 (treatment options identified), MANAGE 2.2 (strategies resourced/implemented), MANAGE 2.3 (emergency interventions defined), MANAGE 2.4 (benefits preserved) | For each high-priority risk, identify a technical (retrain/constrain/add human review), operational (restrict use case), contractual (indemnification), or avoidance (decommission) treatment; define a kill-switch procedure for safety-affecting systems; document benefit-risk tradeoffs for accepted risk |
| MANAGE 3 | AI risk responses are monitored and adjusted; incident response is in place | MANAGE 3.1 (treatment effectiveness monitored), MANAGE 3.2 (incidents documented/investigated), MANAGE 3.3 (lessons applied), MANAGE 3.4 (stakeholders notified) | Implement an AI incident log with severity classification (low/medium/high/critical); define notification thresholds (internal escalation, customer notice, regulatory disclosure); run post-incident reviews that update the risk register and GOVERN policies |
| MANAGE 4 | Risk treatment outcomes are reviewed; lessons learned feed back into GOVERN | MANAGE 4.1 (process effectiveness reviewed), MANAGE 4.2 (improvements implemented), MANAGE 4.3 (lessons update policy), MANAGE 4.4 (risk profile reviewed on major change) | Schedule quarterly AI risk programme reviews across all four functions; use external/third-party audit every 1–2 years; update GOVERN policies and MAP context documents after every major incident or model update |
For the full subcategory list and Playbook-style suggested actions, read references/rmf-core.md.
采取行动解决AI风险并实现AI收益，形成闭环反馈至GOVERN。
| 类别 | 重点 | 代表性子类别 | 具体组织活动 |
|---|---|---|---|
| MANAGE 1 | 对风险进行优先级排序并记录处置措施 | MANAGE 1.1（登记册条目优先级排序/分配）、MANAGE 1.2（反映风险容忍度）、MANAGE 1.3（管理层接受剩余风险） | 为每个风险登记册条目分配处置负责人、目标日期和处置方法；要求高层批准超出容忍度的剩余风险；每年审查剩余风险接受情况 |
| MANAGE 2 | 规划、配置资源并执行AI风险应对策略 | MANAGE 2.1（识别处置选项）、MANAGE 2.2（配置资源并实施策略）、MANAGE 2.3（定义紧急干预措施）、MANAGE 2.4（保留收益） | 针对每个高优先级风险，识别技术（重新训练/约束/添加人工审查）、运营（限制使用场景）、合同（赔偿）或规避（退役）处置措施；为影响安全的系统定义终止开关流程；记录接受风险的收益-风险权衡 |
| MANAGE 3 | 监控并调整AI风险响应；建立事件响应机制 | MANAGE 3.1（监控处置有效性）、MANAGE 3.2（记录/调查事件）、MANAGE 3.3（应用经验教训）、MANAGE 3.4（通知利益相关者） | 实施包含严重性分类（低/中/高/关键）的AI事件日志；定义通知阈值（内部升级、客户通知、监管披露）；开展事后审查，更新风险登记册和GOVERN政策 |
| MANAGE 4 | 审查风险处置结果；经验教训反馈至GOVERN | MANAGE 4.1（审查流程有效性）、MANAGE 4.2（实施改进）、MANAGE 4.3（更新政策）、MANAGE 4.4（重大变更时审查风险画像） | 每季度安排跨四大职能的AI风险项目审查；每1-2年开展外部/第三方审计；在每次重大事件或模型更新后更新GOVERN政策和MAP上下文文档 |
如需完整的子类别列表和操作手册式建议行动，请阅读references/rmf-core.md。
The Seven Trustworthiness Characteristics
七大可信性特征
The AI RMF defines seven characteristics of trustworthy AI. No system is perfectly trustworthy on every dimension — the goal is to make deliberate, documented tradeoffs appropriate to context and risk tolerance. Use the assessment questions below when scoring an AI system or drafting a MEASURE 2 evaluation report.
| Characteristic | Assessment Questions |
|---|---|
| Valid & Reliable | Has the system been tested against its intended use? Does it perform consistently within defined operational limits and across the range of expected conditions? What is out-of-distribution performance? |
| Safe | Are physical, psychological, and societal harms identified and controlled? Is there a defined emergency stop / kill-switch procedure? Have red-team or adversarial exercises estimated real-world failure rates? |
| Secure & Resilient | Is the system hardened against evasion, poisoning, and model extraction/inversion attacks? For LLMs, is it tested against prompt injection? Can it withstand and recover from adversarial or unexpected inputs? |
| Accountable & Transparent | Can decisions be explained and traced to responsible parties? Are roles and responsibilities documented across the lifecycle (GOVERN 2/3)? Is there a model/system card describing purpose, data, and limitations? |
| Explainable & Interpretable | Can the model's behavior be understood by both technical and non-technical audiences? Are SHAP, LIME, counterfactual explanations, or saliency maps available for high-stakes individual decisions? |
| Privacy-Enhanced | Is PII minimized, protected, and handled per applicable law? Are techniques such as differential privacy, k-anonymity, or federated learning applied where appropriate? Is the system resistant to membership-inference attacks? |
| Fair with Harmful Bias Managed | Are demographic biases identified, measured, and mitigated? Is disaggregated performance reported by subgroup? Does disparate impact ratio meet the applicable threshold (e.g., the EEOC "4/5ths rule")? |
For metrics and technical indicators mapped to each characteristic (precision/recall, demographic parity, SHAP/LIME, adversarial accuracy, differential privacy ε, etc.), read references/rmf-profiles.md.
AI RMF定义了可信AI的七大特征。没有系统能在所有维度上完全可信——目标是根据上下文和风险容忍度做出明确、有记录的权衡。在评估AI系统或起草MEASURE 2评估报告时，使用以下评估问题。
| 特征 | 评估问题 |
|---|---|
| 有效且可靠 | 系统是否针对预期用途进行了测试？它是否在定义的操作限制和预期条件范围内持续稳定运行？分布外性能如何？ |
| 安全 | 是否识别并控制了物理、心理和社会危害？是否定义了紧急停止/终止开关流程？是否通过红队或对抗性演练估算了实际失效概率？ |
| 安全且具韧性 | 系统是否能抵御规避、投毒和模型提取/反转攻击？对于大语言模型（LLM），是否针对提示注入进行了测试？它能否承受并从对抗性或意外输入中恢复？ |
| 可问责且透明 | 决策能否被解释并追溯到责任方？全生命周期的角色和职责是否有记录（GOVERN 2/3）？是否有描述用途、数据和局限性的模型/系统卡片？ |
| 可解释且可理解 | 模型行为能否被技术和非技术受众理解？对于高风险个人决策，是否提供SHAP、LIME、反事实解释或显著性图？ |
| 隐私增强 | 是否根据适用法律最小化、保护和处理个人身份信息（PII）？是否在适当情况下应用差分隐私、k-匿名或联邦学习等技术？系统能否抵御成员推断攻击？ |
| 公平且有害偏差可控 | 是否识别、测量并缓解了人口偏差？是否按子群体报告分类性能？差异影响比率是否符合适用阈值（例如：EEOC的“五分之四规则”）？ |
如需与每个特征对应的指标和技术指标（精确率/召回率、人口均等性、SHAP/LIME、对抗性准确性、差分隐私ε等），请阅读references/rmf-profiles.md。
AI Risk Register Template
AI风险登记册模板
Use this column structure for every AI risk register, whether for a single system or an organization-wide inventory. It is deliberately aligned to MAP (identification), MEASURE (TEVV), and MANAGE (treatment) so entries trace cleanly to framework categories.
| Column | Purpose |
|---|---|
| AI System | Name/ID of the AI system or model version |
| Lifecycle Stage | Design / Development / Testing / Deployment / Monitoring / Decommission |
| TEVV Activity | The Test, Evaluation, Verification, or Validation activity that surfaced or measures the risk (e.g., "disaggregated bias testing," "adversarial robustness test") |
| Characteristic at Risk | Which of the seven trustworthiness characteristics is implicated |
| Likelihood / Impact | Qualitative or scored estimate (e.g., Low/Med/High or severity × breadth × reversibility per MAP 4.1) |
| Treatment | Mitigate / Transfer / Avoid / Accept, plus the specific action (MANAGE 2.1) |
| Owner | Individual or role accountable for treatment and residual-risk acceptance |
Worked example row:
| AI System | Lifecycle Stage | TEVV Activity | Characteristic at Risk | Likelihood / Impact | Treatment | Owner |
|---|---|---|---|---|---|---|
| Resume Screening Model v3 | Deployment | Disaggregated performance testing by demographic subgroup (MEASURE 2.2) | Fair with Harmful Bias Managed | High likelihood / High impact — disparate impact ratio measured at 0.71, below the 4/5ths threshold | Mitigate — retrain with rebalanced training data and add human review gate for all rejections in affected subgroup; re-test before re-enabling automated decisions | Head of Talent Acquisition (treatment); Chief AI Officer (residual risk acceptance) |
Second worked example row — third-party model dependency (always include one in register answers):
| AI System | Lifecycle Stage | TEVV Activity | Characteristic at Risk | Likelihood / Impact | Treatment | Owner |
|---|---|---|---|---|---|---|
| Credit Scoring Model (vendor-hosted) | Deployment / Monitoring | Vendor validation-report review + independent benchmark against internal champion model (GOVERN 6.1, MEASURE 2.5) | Valid & Reliable; Accountable & Transparent | Medium likelihood / High impact — vendor retrains without notice; population drift undetected between reviews | Mitigate — contractual change-notification and audit rights, quarterly champion–challenger comparison, documented fallback to prior model version | VP Model Risk (treatment); CRO (residual risk acceptance) |
For financial-services deployments, align treatments with model risk management practice (SR 11-7 / OCC 2011-12): independent validation before use, champion–challenger monitoring in production, effective challenge documented at the model risk committee.
Add rows for every MAP-identified risk; update the Likelihood/Impact and Treatment columns whenever MEASURE produces new evidence (MEASURE 4.3), and close the loop by logging outcomes back to MANAGE 4.
为单个系统或全组织清单创建AI风险登记册时，使用以下列结构。它特意与MAP（识别）、MEASURE（TEVV）和MANAGE（处置）对齐，因此条目可清晰追溯到框架类别。
| 列 | 用途 |
|---|---|
| AI系统 | AI系统或模型版本的名称/ID |
| 生命周期阶段 | 设计 / 开发 / 测试 / 部署 / 监控 / 退役 |
| TEVV活动 | 发现或测量风险的测试、评估、验证或确认活动（例如：“分类偏差测试”、“对抗鲁棒性测试”） |
| 受影响的可信性特征 | 涉及七大可信性特征中的哪一项 |
| 可能性/影响 | 定性或打分估算（例如：低/中/高，或按MAP 4.1的严重性×范围×可逆性打分） |
| 处置措施 | 缓解 / 转移 / 规避 / 接受，加上具体行动（MANAGE 2.1） |
| 负责人 | 负责处置和接受剩余风险的个人或角色 |
示例行：
| AI系统 | 生命周期阶段 | TEVV活动 | 受影响的可信性特征 | 可能性/影响 | 处置措施 | 负责人 |
|---|---|---|---|---|---|---|
| 简历筛选模型v3 | 部署 | 按人口子群体开展分类性能测试（MEASURE 2.2） | 公平且有害偏差可控 | 高可能性/高影响——差异影响比率测得为0.71，低于五分之四阈值 | 缓解——使用重新平衡的训练数据重新训练，并为受影响子群体的所有拒绝决策添加人工审查关卡；重新测试后再启用自动化决策 | 人才招聘主管（处置）；首席AI官（剩余风险接受） |
第二个示例行——第三方模型依赖（登记册答案中必须包含一行）：
| AI系统 | 生命周期阶段 | TEVV活动 | 受影响的可信性特征 | 可能性/影响 | 处置措施 | 负责人 |
|---|---|---|---|---|---|---|
| 信用评分模型（供应商托管） | 部署/监控 | 供应商验证报告审查 + 与内部基准模型的独立对比（GOVERN 6.1，MEASURE 2.5） | 有效且可靠；可问责且透明 | 中可能性/高影响——供应商未经通知重新训练模型；两次审查间未检测到群体漂移 | 缓解——修改合同以获取变更通知和审计权，每季度开展双模型对比，记录回退到旧版本模型的流程 | 模型风险副总裁（处置）；首席风险官（剩余风险接受） |
对于金融服务部署，处置措施需与**模型风险管理实践（SR 11-7 / OCC 2011-12）**对齐：使用前独立验证、生产环境双模型监控、模型风险委员会记录的有效质疑。
为每个MAP识别的风险添加行；每当MEASURE产生新证据时更新“可能性/影响”和“处置措施”列（MEASURE 4.3），并通过将结果记录回MANAGE 4形成闭环。
Common Workflows
常见工作流程
1. GOVERN Gap Assessment
1. GOVERN差距评估
- For each of the 6 GOVERN categories (and their subcategories where granularity is needed), rate status: 🔴 Not Started / 🟡 Partial / 🟢 Implemented
- For each 🔴/🟡, identify the specific gap and the evidence needed to close it (policy document, RACI chart, escalation procedure, etc.)
- Produce a prioritized remediation roadmap (Quick Wins → Medium Term → Long Term), noting that GOVERN gaps typically block progress in MAP/MEASURE/MANAGE
- Flag whether GOVERN is "complete on paper but not operationalized" — a common gap pattern where policies exist but aren't reflected in day-to-day MAP/MEASURE/MANAGE activity (see references/rmf-profiles.md)
- Deliver these mini-templates in the answer itself (fill them with the organization's specifics):
- AI risk policy outline (GOVERN 1.2): 1. Purpose & scope · 2. Definitions & AI system inventory criteria · 3. Risk tolerance statement · 4. Roles & accountability (RACI) · 5. Lifecycle requirements (MAP/MEASURE/MANAGE gates per stage) · 6. Third-party AI requirements · 7. Incident response & escalation · 8. Review cadence
- AI governance committee charter (RACI skeleton) (GOVERN 2.1): Accountable — executive sponsor/CAIO; Responsible — AI product owners, data science leads; Consulted — legal, privacy, security, HR for employment uses; Informed — audit, board risk committee; quorum, meeting cadence, decision rights (approve/deny deployment, accept residual risk)
- Risk tolerance statement examples (GOVERN 1.3): "We do not deploy AI that makes fully automated adverse decisions about individuals without human review"; "Disaggregated performance gaps above X% between demographic groups block deployment until remediated"
- AI inventory minimum fields (GOVERN 1.6): system name/owner · purpose & users · model type/provenance (built/bought/fine-tuned) · data categories · lifecycle stage · risk tier · last TEVV date
- 针对6个GOVERN类别（必要时细化到子类别），评估状态：🔴未启动 / 🟡部分完成 / 🟢已实施
- 针对每个🔴/🟡状态，识别具体差距和填补差距所需的证据（政策文档、RACI图表、升级流程等）
- 生成优先级修复路线图（快速赢 → 中期 → 长期），注意GOVERN差距通常会阻碍MAP/MEASURE/MANAGE的进展
- 标记GOVERN是否“纸面完成但未落地”——这是常见的差距模式，即政策存在但未体现在日常MAP/MEASURE/MANAGE活动中（见references/rmf-profiles.md）
- 在答案中直接提供以下迷你模板（填入组织的具体信息）：
- AI风险政策大纲（GOVERN 1.2）：1. 目的与范围 · 2. 定义与AI系统清单标准 · 3. 风险容忍度声明 · 4. 角色与问责（RACI） · 5. 生命周期要求（各阶段的MAP/MEASURE/MANAGE关卡） · 6. 第三方AI要求 · 7. 事件响应与升级 · 8. 审查周期
- AI治理委员会章程（RACI框架）（GOVERN 2.1）：问责方——执行发起人/首席AI官；负责方——AI产品负责人、数据科学主管；咨询方——法务、隐私、安全、HR（针对就业场景）；告知方——审计、董事会风险委员会；法定人数、会议周期、决策权（批准/拒绝部署、接受剩余风险）
- 风险容忍度声明示例（GOVERN 1.3）：“我们不会部署完全自动化做出对个人不利决策且无人工审查的AI”；“人口子群体间的分类性能差距超过X%时，需修复后方可部署”
- AI清单必填字段（GOVERN 1.6）：系统名称/负责人 · 用途与用户 · 模型类型/来源（自研/采购/微调） · 数据类别 · 生命周期阶段 · 风险等级 · 上次TEVV日期
2. Hiring / Employment AI Risk Assessment
2. 招聘/就业AI风险评估
- MAP: Document intended use (MAP 1.2), affected populations — applicants, current employees (MAP 1.4), and prohibited uses (e.g., no fully automated rejection without human review)
- MAP: Build the stakeholder risk/benefit matrix (MAP 3.1) — employer efficiency benefit vs. applicant risk of disparate impact
- MEASURE: Run disaggregated performance and disparate-impact-ratio testing across protected classes (MEASURE 2.2); document explainability approach for adverse decisions (MEASURE 2.3)
- MANAGE: Define treatment for any subgroup below the 4/5ths threshold — retraining, human-in-the-loop review, or use restriction (MANAGE 2.1)
- Populate the AI Risk Register (above) with one row per identified hiring-stage risk
- Cross-reference sector considerations: EEOC, NYC Local Law 144, and EU AI Act high-risk classification for employment AI (see references/rmf-profiles.md)
- MAP：记录预期用途（MAP 1.2）、受影响群体——求职者、现有员工（MAP 1.4），以及禁止用途（例如：无人工审查的完全自动化拒绝）
- MAP：构建利益相关者风险/收益矩阵（MAP 3.1）——雇主效率收益vs求职者面临的差异影响风险
- MEASURE：针对受保护群体开展分类性能和差异影响比率测试（MEASURE 2.2）；记录不利决策的可解释方法（MEASURE 2.3）
- MANAGE：为任何低于五分之四阈值的子群体定义处置措施——重新训练、人工介入审查或限制使用（MANAGE 2.1）
- 使用上述模板填充AI风险登记册，为每个招聘阶段识别的风险添加一行
- 交叉参考行业要求：EEOC、纽约市第144号地方法规，以及《欧盟AI法案》中就业AI的高风险分类（见references/rmf-profiles.md）
3. Credit Scoring Risk Register
3. 信用评分风险登记册
- MAP: Document context of use (MAP 1.2/1.3) — loan origination, limit decisions, pricing — and legal constraints (ECOA, Fair Housing Act, EU AI Act high-risk classification)
- MAP: Identify affected stakeholders and prioritize risks that are irreversible or affect a protected class (MAP 4.1)
- MEASURE: Test fairness metrics (demographic parity, equalized odds, disparate impact ratio) and explainability sufficient to produce adverse-action notices (MEASURE 2.2/2.3)
- MANAGE: Document treatment — model adjustment, threshold changes, or human review escalation — and residual risk acceptance by an accountable officer (MANAGE 1.3)
- Build the risk register using the template above, with "Characteristic at Risk" typically Fair with Harmful Bias Managed or Accountable & Transparent
- MAP：记录使用上下文（MAP 1.2/1.3）——贷款发放、额度决策、定价——以及法律约束（ECOA、公平住房法案、《欧盟AI法案》高风险分类）
- MAP：识别受影响利益相关者，并优先处理不可逆或影响受保护群体的风险（MAP 4.1）
- MEASURE：测试公平性指标（人口均等性、平等机会、差异影响比率），并提供足以生成不利行动通知的可解释性（MEASURE 2.2/2.3）
- MANAGE：记录处置措施——模型调整、阈值变更或人工审查升级——以及负责人对剩余风险的接受（MANAGE 1.3）
- 使用上述模板构建风险登记册，“受影响的可信性特征”通常为公平且有害偏差可控或可问责且透明
4. Incident Response (MANAGE 3)
4. 事件响应（MANAGE 3）
- Trigger conditions: model accuracy degradation, bias threshold breach, adversarial attack, data drift
- Response steps: Contain → Assess impact → Notify stakeholders → Remediate → Document → Update risk register → Feed lessons learned into GOVERN (MANAGE 4.3)
- Classify severity (low/medium/high/critical) and pre-define notification thresholds: internal escalation, customer notice, regulatory disclosure
- 触发条件：模型准确率下降、偏差阈值突破、对抗性攻击、数据漂移
- 响应步骤：遏制 → 评估影响 → 通知利益相关者 → 修复 → 记录 → 更新风险登记册 → 将经验教训反馈至GOVERN（MANAGE 4.3）
- 严重性分类（低/中/高/关键）并预定义通知阈值：内部升级、客户通知、监管披露
AI Risk Profiles
AI风险画像
An AI Risk Profile is an organization's customization of the AI RMF to reflect its specific AI use cases, applicable laws, defined risk tolerance, and the trustworthiness characteristics most relevant to its systems. The AI RMF defines two profile types:
| Profile Type | Description | Use |
|---|---|---|
| Current Profile | Where the organization is today — which categories are implemented and to what degree | Baseline assessment |
| Target Profile | Where the organization wants to be — desired maturity for each category | Gap analysis and roadmap |
The gap between Current and Target Profile drives the risk management roadmap:
- Scope — Define which AI systems are in scope (all AI, specific high-risk systems, or a single system)
- Assess Current State — Rate each of the 19 categories: Not Started (0) / Partial (1) / Implemented (2) / Optimized (3)
- Set Target State — Define desired maturity per category based on risk tolerance and regulatory requirements
- Gap Analysis — Categories where Target > Current are gaps requiring action
- Prioritize — Weight gaps by the risk they represent; address highest-risk gaps first
- Roadmap — Assign owners, timelines, and resources to close each gap
NIST also uses cross-sectoral and use-case profiles as companions to the core AI RMF (for example, a Generative AI Profile addressing risks specific to generative AI systems). When a user's question concerns generative-AI-specific risk, apply the same GOVERN/MAP/MEASURE/MANAGE structure and trustworthiness characteristics above, and note explicitly that generative-AI-specific subcategory detail should be verified against the current NIST publication rather than assumed.
AI风险画像是组织根据自身特定AI用例、适用法律、定义的风险容忍度以及与系统最相关的可信性特征，对AI RMF进行的定制。AI RMF定义了两种画像类型：
| 画像类型 | 描述 | 用途 |
|---|---|---|
| 当前画像 | 组织当前的状态——哪些类别已实施及实施程度 | 基线评估 |
| 目标画像 | 组织期望达到的状态——每个类别的期望成熟度 | 差距分析和路线图 |
当前画像与目标画像之间的差距驱动风险管理路线图：
- 范围——定义纳入范围的AI系统（所有AI、特定高风险系统或单个系统）
- 评估当前状态——对19个类别进行评级：未启动（0）/部分完成（1）/已实施（2）/优化（3）
- 设定目标状态——根据风险容忍度和监管要求定义每个类别的期望成熟度
- 差距分析——目标状态>当前状态的类别即为需要行动的差距
- 优先级排序——根据风险权重对差距排序；优先处理最高风险的差距
- 路线图——为每个差距分配负责人、时间表和资源
NIST还提供跨行业和用例的画像作为核心AI RMF的补充（例如：针对生成式AI系统特定风险的生成式AI画像）。当用户的问题涉及生成式AI特定风险时，应用上述GOVERN/MAP/MEASURE/MANAGE结构和可信性特征，并明确指出生成式AI特定子类别细节需参考当前NIST出版物，而非假设。
Cross-Framework Mapping
跨框架映射
NIST AI RMF ↔ EU AI Act (Regulation (EU) 2024/1689)
NIST AI RMF ↔ 《欧盟AI法案》（Regulation (EU) 2024/1689）
| AI RMF Function | EU AI Act Requirement |
|---|---|
| GOVERN 1 (AI risk policies) | Art. 9 (Risk management system) for high-risk AI |
| GOVERN 2/3 (Accountability) | Art. 16 (Obligations of high-risk AI providers), Art. 26 (Deployer obligations) |
| MAP 1 (Intended use) | Art. 9(2) — risk management must cover intended and reasonably foreseeable misuse |
| MAP 3 (Stakeholder mapping) | Art. 9(2)(b) — identification and analysis of known and foreseeable risks |
| MEASURE 2 (System evaluation) | Art. 10 (Data governance), Art. 15 (Accuracy, robustness, cybersecurity) |
| MEASURE 3 (Ongoing monitoring) | Art. 72 (Post-market monitoring), Art. 26(5) — deployer monitoring obligations |
| MANAGE 3 (Incident response) | Art. 73 (Reporting of serious incidents to market surveillance) |
| All functions | Annex IX (Technical documentation requirements for high-risk AI systems) |
Key difference: The EU AI Act is mandatory for in-scope providers and deployers; the NIST AI RMF is voluntary. Organizations subject to the EU AI Act should use the NIST AI RMF as the risk management methodology that satisfies Art. 9's "appropriate risk management system" requirement.
| AI RMF职能 | 《欧盟AI法案》要求 |
|---|---|
| GOVERN 1（AI风险政策） | 第9条（风险管理体系）适用于高风险AI |
| GOVERN 2/3（问责） | 第16条（高风险AI提供商义务）、第26条（部署方义务） |
| MAP 1（预期用途） | 第9(2)条——风险管理必须覆盖预期和可合理预见的误用 |
| MAP 3（利益相关者映射） | 第9(2)(b)条——识别和分析已知及可预见的风险 |
| MEASURE 2（系统评估） | 第10条（数据治理）、第15条（准确性、鲁棒性、网络安全） |
| MEASURE 3（持续监控） | 第72条（上市后监控）、第26(5)条——部署方监控义务 |
| MANAGE 3（事件响应） | 第73条（向市场监管机构报告严重事件） |
| 所有职能 | 附件IX（高风险AI系统的技术文档要求） |
关键差异：《欧盟AI法案》对纳入范围的提供商和部署方具有强制性；而NIST AI RMF是自愿性的。受《欧盟AI法案》约束的组织可将NIST AI RMF作为满足第9条“适当风险管理体系”要求的方法。
NIST AI RMF ↔ ISO/IEC 42001:2023
NIST AI RMF ↔ ISO/IEC 42001:2023
| AI RMF Function/Category | ISO 42001 Equivalent |
|---|---|
| GOVERN 1 (Policies in place) | Clause 5 (Leadership), Clause 6 (Planning), A.2 (AI policy) |
| GOVERN 2 (Accountability) | Clause 5.3 (Roles and responsibilities), A.2.3 |
| GOVERN 3 (Roles) | Clause 5.3, A.2.5 (Responsibilities for AI system impact) |
| GOVERN 4 (Cross-functional teams) | Clause 7.1 (Resources), A.2.5 |
| GOVERN 5 (Risk tolerance) | Clause 6.1 (Risk and opportunity), A.5.2 (AI risk assessment) |
| MAP 1 (Context) | Clause 4 (Context of organization), A.3 (Internal/external context) |
| MAP 2 (Scientific understanding) | A.6 (AI system lifecycle) |
| MAP 3 (Stakeholder risk/benefit) | Clause 4.2 (Interested parties), A.8.4 (Impact assessment) |
| MAP 5 (Likelihood/impact) | A.5.2 (AI risk assessment methodology) |
| MEASURE 2 (System evaluation) | A.6.2 (AI system design), A.10 (Use of AI systems) |
| MEASURE 3 (Ongoing monitoring) | Clause 9.1 (Monitoring and measurement), A.6.2.5 |
| MANAGE 2 (Treatment strategies) | Clause 6.1.3 (AI risk treatment), A.5.3 |
| MANAGE 3 (Incident response) | A.9 (Performance evaluation), Clause 10 (Improvement) |
| MANAGE 4 (Review and improve) | Clause 10.2 (Nonconformity), Clause 9.3 (Management review) |
For NIST CSF 2.0 and NIST Privacy Framework mappings, sector-specific risk considerations (healthcare, financial services, HR/recruitment, criminal justice, government, education, autonomous systems), implementation tiers, and common gap patterns, read references/rmf-profiles.md.
| AI RMF职能/类别 | ISO 42001对应条款 |
|---|---|
| GOVERN 1（建立政策） | 第5条（领导力）、第6条（规划）、A.2（AI政策） |
| GOVERN 2（问责） | 第5.3条（角色与职责）、A.2.3 |
| GOVERN 3（角色） | 第5.3条、A.2.5（AI系统影响的职责） |
| GOVERN 4（跨职能团队） | 第7.1条（资源）、A.2.5 |
| GOVERN 5（风险容忍度） | 第6.1条（风险与机遇）、A.5.2（AI风险评估） |
| MAP 1（上下文） | 第4条（组织上下文）、A.3（内部/外部上下文） |
| MAP 2（科学认知） | A.6（AI系统生命周期） |
| MAP 3（利益相关者风险/收益） | 第4.2条（利益相关方）、A.8.4（影响评估） |
| MAP 5（可能性/影响） | A.5.2（AI风险评估方法） |
| MEASURE 2（系统评估） | A.6.2（AI系统设计）、A.10（AI系统使用） |
| MEASURE 3（持续监控） | 第9.1条（监控与测量）、A.6.2.5 |
| MANAGE 2（处置策略） | 第6.1.3条（AI风险处置）、A.5.3 |
| MANAGE 3（事件响应） | A.9（绩效评估）、第10条（改进） |
| MANAGE 4（审查与改进） | 第10.2条（不符合项）、第9.3条（管理评审） |
如需NIST CSF 2.0和NIST隐私框架映射、行业特定风险考量（医疗、金融服务、HR/招聘、司法、政府、教育、自治系统）、实施层级和常见差距模式，请阅读references/rmf-profiles.md。
Reference Files
参考文件
For deeper content, read these files as needed:
- references/rmf-core.md — All 19 categories with full subcategory descriptions and Playbook-style suggested actions for GOVERN, MAP, MEASURE, and MANAGE
- references/rmf-profiles.md — AI Risk Profiles, trustworthy AI metrics and indicators, sector-specific guidance, cross-framework mapping (ISO 42001, EU AI Act, NIST CSF, NIST Privacy Framework), implementation tiers, and common gap patterns
This skill provides general compliance information, not legal advice. Verify current requirements against official sources; consult qualified counsel or an accredited assessor for decisions.
如需更深入的内容，可根据需要阅读以下文件：
- references/rmf-core.md——所有19个类别，包含GOVERN、MAP、MEASURE和MANAGE的完整子类别描述及操作手册式建议行动
- references/rmf-profiles.md——AI风险画像、可信AI指标和技术指标、行业特定指南、跨框架映射（ISO 42001、《欧盟AI法案》、NIST CSF、NIST隐私框架）、实施层级和常见差距模式
本技能提供一般性合规信息，而非法律建议。请对照官方来源验证当前要求；决策时请咨询合格律师或认证评估师。