AI GRCA Practice Lab

How should NIST AI RMF best be described to a business stakeholder?

A voluntary US framework that structures how organisations identify and manage AI riskA certification scheme operated by ISOA replacement for all national AI legislationA binding international law that all AI systems must certify against

A committee is deciding who can approve exceptions to the AI policy and how model ownership is assigned. Which NIST AI RMF function does this belong to?

MeasureGovernManageMap

A team documents the intended use, foreseeable misuse, affected people and deployment environment of a new claims-triage system before building controls. This work sits primarily in:

ManageGovernMapMeasure

A red-teaming exercise probes a chatbot for prompt-injection vulnerabilities and records the pass/fail results against a threshold. This activity sits primarily in:

ManageGovernMapMeasure

After testing reveals a high residual risk, the organisation decides to add a mandatory human review step and sets a monitoring threshold for post-launch drift. This decision sits primarily in:

ManageGovernMapMeasure

Which OECD AI Principle is most directly reflected by publishing a model card with intended use, limitations and evaluation results?

Inclusive growth, sustainable development and well-beingTransparency and explainabilityRobustness, security and safety onlyAccountability only, with no relevance to transparency

A senior leader says 'we don't need NIST, ISO or OECD — we only care about the law.' What is the most accurate response?

ISO 42001 replaces the need for any legal analysisThey are right; voluntary frameworks are irrelevant to complianceVoluntary frameworks like NIST and OECD often operationalise the practices that legal obligations (e.g. EU AI Act risk management, UK GDPR accountability) actually require in practiceNIST AI RMF is itself a binding EU regulation

A company sets its AI risk appetite as 'no automated decision may adversely affect a customer's access to credit without a documented human review path.' This statement is best described as defining:

An audit findingA lawful basisA control activityRisk appetite

In the MAP function, which of the following is a 'foreseeable misuse' scenario for a public-facing HR chatbot, as opposed to intended use?

A jobseeker trying to extract another candidate's personal data through crafted promptsAn HR admin updating the chatbot's FAQ contentA manager reading the chatbot's usage dashboardAn employee asking the chatbot to summarise the leave policy

Which is the strongest example of a MEASURE-stage acceptance criterion?

'The model should perform well across groups''False-negative rate for high-severity claims must not exceed 2%, tested on a held-out sample, with any breach blocking release''The supplier says the model is safe''We will decide what counts as good enough after launch'

A model's accuracy is degrading over time as real-world claims patterns shift away from the training data. This phenomenon is best described as:

Purpose limitation failureData poisoningDriftPrompt injection

During MANAGE, a residual risk remains 'high' even after controls are applied. Who should make the final acceptance decision?

The affected customers, informallyThe data scientist who built the modelWhoever is fastest to respond before launchAn authorised risk-acceptance role at a level matched to the risk's severity

Which best distinguishes a 'profile' from the 'core' in NIST AI RMF?

The core defines the four functions and categories; a profile tailors their application to a specific use case, sector or risk toleranceProfiles apply only to generative AI; the core applies only to classic AIThere is no meaningful differenceA profile is a mandatory certification; the core is optional guidance

An organisation wants to demonstrate 'accountability' in the OECD sense for an automated loan-decision tool. Which action best supports this?

Removing all logs to reduce storage costAssigning a named accountable owner and maintaining records that allow the decision-making process to be reconstructed and challengedLetting the model decide who is accountableAvoiding any human sign-off to speed up processing

A benchmark shows a model performs well on a public test set but the deployment team has not tested it on the organisation's real, messier data. What is the primary governance concern?

Benchmarks eliminate the need for acceptance criteriaNone — public benchmarks are always representative of production useBenchmark performance may not generalise to the organisation's actual data and context, so it cannot substitute for context-specific testingPublic benchmarks are illegal to rely on

Which statement best reflects the relationship between NIST AI RMF and ISO/IEC 42001?

ISO/IEC 42001 replaces the need for MAP and MEASURE entirelyNIST AI RMF is a certifiable management-system standard like ISO/IEC 42001They are unrelated and cannot be used togetherNIST AI RMF offers risk-management functions that can inform the risk-related clauses of an ISO/IEC 42001 AI management system

A supplier changes the underlying model behind an API with no advance notice, and downstream accuracy drops. Which NIST AI RMF weakness does this best illustrate?

A MAP failure to consider supplier dependencies and change riskA MEASURE failure because testing is impossible for hosted modelsA Govern strength because the supplier is accountableNot a governance issue because the contract is commercial, not technical

Which of these best reflects OECD's 'human-centred values and fairness' principle in a real deployment?

Maximising throughput regardless of individual impactDesigning the system so it respects human rights, democratic values and fair treatment across affected groupsRemoving all human review to reduce bias from staffOnly applying to systems sold in OECD member countries

A team wants to reuse the same MEASURE test plan for every AI system regardless of risk level or context. What is the main problem with this approach?

Reusing test plans violates NIST AI RMF by definitionThere is no problem; standard test plans are always sufficientAcceptance criteria and test coverage should be proportionate to the risks identified in MAP, not identical across all use casesMEASURE should only be done once per organisation, not per system

A company documents a full risk register but has no evidence that mitigations were actually implemented or that anyone checks whether they still work. Which MANAGE-related weakness does this show?

This is a Map-stage problem onlyThere is no weakness; documentation alone is sufficient assuranceRisk identification is missingRisk treatment exists on paper but lacks implementation and monitoring evidence

Checkpoint 2

How should NIST AI RMF best be described to a business stakeholder?

A committee is deciding who can approve exceptions to the AI policy and how model ownership is assigned. Which NIST AI RMF function does this belong to?

A team documents the intended use, foreseeable misuse, affected people and deployment environment of a new claims-triage system before building controls. This work sits primarily in:

A red-teaming exercise probes a chatbot for prompt-injection vulnerabilities and records the pass/fail results against a threshold. This activity sits primarily in:

After testing reveals a high residual risk, the organisation decides to add a mandatory human review step and sets a monitoring threshold for post-launch drift. This decision sits primarily in:

Which OECD AI Principle is most directly reflected by publishing a model card with intended use, limitations and evaluation results?

A senior leader says 'we don't need NIST, ISO or OECD — we only care about the law.' What is the most accurate response?

A company sets its AI risk appetite as 'no automated decision may adversely affect a customer's access to credit without a documented human review path.' This statement is best described as defining:

In the MAP function, which of the following is a 'foreseeable misuse' scenario for a public-facing HR chatbot, as opposed to intended use?

Which is the strongest example of a MEASURE-stage acceptance criterion?

A model's accuracy is degrading over time as real-world claims patterns shift away from the training data. This phenomenon is best described as:

During MANAGE, a residual risk remains 'high' even after controls are applied. Who should make the final acceptance decision?

Which best distinguishes a 'profile' from the 'core' in NIST AI RMF?

An organisation wants to demonstrate 'accountability' in the OECD sense for an automated loan-decision tool. Which action best supports this?

A benchmark shows a model performs well on a public test set but the deployment team has not tested it on the organisation's real, messier data. What is the primary governance concern?

Which statement best reflects the relationship between NIST AI RMF and ISO/IEC 42001?

A supplier changes the underlying model behind an API with no advance notice, and downstream accuracy drops. Which NIST AI RMF weakness does this best illustrate?

Which of these best reflects OECD's 'human-centred values and fairness' principle in a real deployment?

A team wants to reuse the same MEASURE test plan for every AI system regardless of risk level or context. What is the main problem with this approach?

A company documents a full risk register but has no evidence that mitigations were actually implemented or that anyone checks whether they still work. Which MANAGE-related weakness does this show?