ISACA Advanced in AI Risk (AAIR) (AAIR) (AAIR) — Questions 76–150

An AI model inventory exposes every model in use, including unsanctioned ones deployed outside governance, directly satisfying the need to identify and mitigate shadow AI risks. By recording ownership, purpose and deployment status, it surfaces unapproved models that Microsoft Entra ID controls and monitoring would otherwise miss, enabling remediation.

Suspending the model immediately contains the bias exposure while a formal root cause analysis traces the training-data defect, satisfying the Governance function's mandate to enforce accountability and risk controls before remediation. Governance prioritises containment and documented investigation over technical fixes, which belong to development teams once the risk is bounded.

Unauthorised access to the model environment is an infrastructure risk because it targets the hosting, credentials and network perimeter rather than the model's logic. Compromise enables tampering, data exfiltration or denial of service, satisfying the stem's infrastructure category.

Time to complete a risk assessment for new models directly quantifies process throughput, the core of efficiency. It measures how quickly the risk function handles each new model, exposing bottlenecks and manual effort, unlike coverage or accuracy metrics that gauge effectiveness rather than speed of delivery.

Shadow AI covers unsanctioned tools and models deployed outside IT visibility, so its primary risk is bypassing established AI risk governance and compliance controls. Because these systems evade approved review, monitoring and data-handling requirements, the organisation loses assurance over regulatory and ethical obligations.

KL divergence quantifies how far inference feature distributions have shifted from training data, detecting covariate drift without labels. Since ground truth is unavailable, this unsupervised measure flags degradation by exposing input-distribution changes that undermine model assumptions, satisfying the no-real-time-labels constraint directly.

Override data is a valuable feedback signal. Systematically analyzing when and why clinicians deviate from recommendations can reveal miscalibrated thresholds, subgroup performance gaps, or workflow mismatches. Recalibrating the model and tuning alert thresholds addresses the root cause while preserving clinician discretion, which is the appropriate mitigation for a decision support tool.

When the training distribution shifts on every retrain, the learned parameters and resulting predictions change unpredictably, so identical inputs no longer yield reproducible outputs. This destroys auditability and makes isolating whether a fault lies in data or model effectively impossible.

Aligning the AI RMF functions with established ERM activities allows AI risks to be identified, assessed, treated, and monitored through processes the insurer already operates. This produces traceability from framework outcomes to risk register entries and board reporting, which is exactly the integration the CRO requires.

Access management controls restrict who and what can invoke the model, enforcing least privilege and preventing unauthorised use or data exposure. This satisfies the risk-mitigation requirement by governing model consumption, complementing data, monitoring and lifecycle controls within the Control Framework.

Recording the training dataset version ties the model to the precise data it learned from, enabling auditors to trace bias, provenance and reproducibility. Without it, the model's behaviour cannot be reconstructed or defended during an audit review.

Developing a board-approved AI risk appetite statement directly supports the integration of risk appetite into decision-making. It provides clear guidance on acceptable risk levels, enabling consistent risk-based decisions across the organization. The AI RMF Govern function emphasizes the importance of a risk appetite framework. Board approval ensures top-level commitment and authority, making it a cornerstone of effective AI risk governance.

Pre-trained base models encode statistical patterns from their original training corpora, including societal biases that cannot be traced to individual records or removed by fine-tuning. Transfer learning propagates these embedded biases into the downstream AI model, so the critical risk is inherited, unidentifiable bias requiring provenance scrutiny and bias evaluation before reuse.

The Compliance/Legal Department and the Data Protection Officer hold statutory accountability for GDPR adherence, interpreting lawful bases, data subject rights and cross-border transfer rules. They independently audit AI models and enforce remediation, satisfying the stem's requirement for regulatory compliance oversight rather than technical model development or deployment.

Chief Legal Counsel supplies legal and regulatory expertise on AI liability, privacy and compliance obligations, which no purely technical or business role covers. This ensures the committee's holistic risk view demanded by the stem, spanning legal exposure alongside operational and ethical concerns.

The inventory is the central register recording each model's metadata, risk rating and accountable owner, enabling ongoing oversight, review and audit evidence. This directly satisfies the governance requirement to maintain traceability and accountability across the model lifecycle.

The model was validated only on data resembling existing channels, so its operating domain was never bounded to include the new mobile channel. Documenting intended use and domain boundaries during design would have flagged that mobile claims are out of scope, triggering revalidation before launch. This is the lifecycle control whose absence directly caused the gap.

The Manage function is responsible for allocating resources to address AI risks, implementing risk responses, and continuously monitoring risks and controls. It involves ongoing oversight, updating controls as needed, and ensuring that risk management is dynamic. This directly addresses the need for continuous monitoring and updating of controls. Manage ensures that risk treatments are effective and adjusted over time.

The AI Risk Appetite Statement is the governing document that formally records how much AI-related risk the organisation is willing to accept, giving risk owners the thresholds against which AI initiatives and controls are assessed.

Effective AI risk governance requires clearly assigned accountability for model outcomes. Having no designated owner for production results means no one is answerable for monitoring, escalation or remediation, which is a direct governance failure rather than a technical or data-quality issue.

Evaluating likelihood and impact quantifies each identified threat, producing the risk score that drives prioritisation. This analytical step is core to AI risk assessment, satisfying the scenario's requirement for a task within that process rather than threat identification or control implementation.

An AI Risk Assessment Protocol or template provides the structured methodology, scoring criteria, and documentation fields needed to begin assessing the machine learning project consistently. It establishes the repeatable process before any project-specific analysis is performed.

Full lineage records each dataset, transformation and version feeding every training run, so corrupted or drifted inputs can be traced to their source and excluded. This directly satisfies the stem's data quality risk constraint by making provenance auditable rather than assumed.

Removing the model from production service halts the harmful outputs that breach the Risk Appetite threshold, satisfying the requirement to stop ongoing risk exposure. Remediation then addresses the underlying drift or data issue before redeployment, rather than merely alerting or adjusting the threshold.

Risk Treatment/Mitigation is a core phase of the AI Risk Management lifecycle, where identified and assessed risks are addressed through controls, transfer, avoidance or acceptance. It follows risk identification and analysis, reducing residual risk to acceptable levels before monitoring.

Splitting data into train, validation and test sets lets practitioners measure performance on unseen data, revealing whether the model generalises or merely memorises training examples. This directly mitigates overfitting risk and supports unbiased model selection before deployment.

The Chief Risk Officer owns the enterprise risk appetite and reporting to the board, so chairing the AI Risk Governance Committee ensures AI risks are assessed against that same appetite. This satisfies the requirement for enterprise-wide alignment rather than a technology-focused or business-unit perspective.

Automated performance validation against a hold-out test set applies defined threshold gates, providing an objective, reproducible check before promotion. This satisfies the stem's requirement for a final gate, since it blocks models failing accuracy, fairness or latency thresholds rather than relying on manual judgement.

Boards must weigh AI investment against ethical, regulatory and operational exposure, so literacy gives directors the vocabulary to interrogate risk appetite, challenge assurance evidence and approve adoption decisions with informed oversight rather than deferring entirely to technical teams.

An enterprise AI inventory catalogues every deployed model, dataset, and owner, giving the visibility needed to assess, prioritise, and monitor risk continuously. Without this baseline, governance controls, impact assessments, and mitigation tracking cannot be scoped or sustained across the organisation.

A risk appetite statement defines the amount and type of AI risk the organisation is willing to accept, giving decision-makers explicit boundaries for judging whether outcomes are acceptable. It directly satisfies the requirement for clear guidance on acceptable AI outcomes within the policy.

Concept drift means the statistical relationship between inputs and target changes, so the model's predictions become less accurate over time. That degradation is a governance concern because decisions based on stale models can cause harm, bias, or compliance breaches requiring oversight.

The Chief Risk Officer owns the enterprise risk appetite and translates it into tolerances. Including them ensures AI initiatives are governed against that appetite, aligning technical capability with organisational risk limits rather than leaving risk oversight to engineering alone.

Design and problem formulation defines intended use, affected stakeholders, and constraints before any data or model exists. Embedding risk controls here prevents harm that later phases can only mitigate. This satisfies the stem's requirement for initial risk mitigation, since downstream phases inherit these foundational decisions.

The AI Governance Charter or Policy formally assigns roles, decision rights, and oversight bodies, making it the foundational reference for the governance structure. Other artefacts such as risk registers or model cards operate within the authority this charter establishes.

Output guardrails and moderation layers inspect model responses before delivery, filtering harmful or inappropriate content that input controls alone cannot catch. This directly satisfies the stem's requirement for the governance control most critical to preventing harmful generative output.

Embedding governance controls into existing GRC and CI/CD tooling removes the need for separate approval portals, so teams satisfy policy checks within pipelines they already use. This directly addresses the adoption constraint: governance that adds friction or duplicate process is bypassed, whereas native integration makes compliance the path of least resistance.

Integrating AI risk into ERM must begin with an inventory: identifying and classifying every AI application by criticality and risk level. Without this baseline, subsequent assessment, treatment and monitoring cannot be scoped or prioritised, so it is the foundational first step.

When model internals are unavailable, behavioral monitoring is the workable control. Tracking input feature drift and output score distributions against the baselines captured during validation provides an early signal of degradation, while periodic reconciliation against actual prior-authorization outcomes confirms whether scores still correspond to real-world results. This satisfies the audit request without requiring weights or training data from the vendor.

Comprehensive audit trails capture model lineage, versions and validation results, giving regulators an immutable record of how each model was built, changed and approved. This directly satisfies the external regulatory requirement for demonstrable, traceable model governance evidence across the model lifecycle.

Fairness audits apply specialised statistical metrics such as Disparate Impact to quantify outcome disparities across protected groups. Measuring bias on defined axes lets teams detect and remediate discriminatory behaviour before deployment, satisfying the mitigation requirement rather than merely documenting intent.

Performance metrics such as accuracy or latency quantify only technical behaviour. They cannot capture social harm, fairness, privacy or regulatory compliance obligations, so relying on them alone leaves those risk dimensions unmanaged. Risk management therefore requires broader assessment beyond model metrics.

A five-year-old dataset no longer reflects current populations, behaviours or conditions, so the model learns stale patterns that misrepresent present reality. This is a data quality risk because representativeness, a core quality dimension, has degraded over time.

Aligning scoring criteria lets AI risks be rated on the same likelihood and impact scales as other enterprise risks, so they aggregate into existing heat maps and reporting rather than sitting in a separate silo. This directly satisfies the stem's requirement to integrate AI risk into the broader ERM framework.

Anchoring tolerance thresholds to human-rights and safety impact, varied by deployment context, ties acceptable risk directly to the NIST AI RMF's harm-based framing. This lets teams move quickly in low-stakes contexts while constraining high-impact deployments, aligning velocity with organisational risk tolerance.

An advisory role keeps ethics guidance feeding into governance without duplicating decision rights. The committee supplies ethical judgements and recommendations, while the governance body retains accountability for risk decisions and controls, satisfying the stem's requirement for a defined interaction between the two bodies.

Mapping AI metrics to business impact indicators directly satisfies the executive buy-in constraint, because leadership allocates budget against revenue, churn and downtime rather than model-centric scores such as precision or latency. Framing risk monitoring in those financial and operational terms makes the dashboard's value legible to decision-makers who lack technical AI backgrounds.

End-to-end lineage and version control logs record every dataset, transformation, and model version used, providing regulators with traceable provenance. This satisfies the regulated industry's audit requirement by evidencing exactly how the model was built and changed over time.

Documented independent MRM validation demonstrates that models face objective scrutiny separate from their developers, directly evidencing the accountability principle. The percentage metric shows governance coverage across the model inventory, proving oversight is systematic rather than ad hoc.

Risk management begins during design and requirements, where intended purpose, affected stakeholders and potential harms are identified before development. Addressing risk here shapes architecture and data decisions, rather than reacting after deployment when remediation is costlier.

Reporting roadmap status and mitigation effectiveness gives the board forward-looking assurance on residual AI risk, satisfying the stem's requirement to communicate risk posture rather than raw operational metrics. It shows which controls are live, which remain outstanding, and whether deployed mitigations actually reduce the identified exposures quarter on quarter.

Hallucination risk means LLMs can generate fluent but fabricated content, so mandating human-in-the-loop review for customer-facing output ensures a person verifies accuracy before publication. This directly addresses the risk appetite element of acceptable AI error tolerance in externally visible content.

AI regulations differ by jurisdiction, so local counsel translates region-specific rules into actionable committee guidance. Their legal interpretation satisfies the global coordination constraint, ensuring the steering committee accounts for each operational territory's distinct compliance obligations rather than applying one uniform standard.

Govern cultivates a risk-management culture by assigning accountability, defining policies, and documenting procedures across the AI lifecycle. It supplies the organisational scaffolding the other functions — Map, Measure, and Manage — rely on, rather than performing technical measurement itself.

Independent assurance is the crux: validation is performed by parties separate from the development team, so it satisfies governance's need for objective evidence that the model stays within defined risk thresholds. Testing, by contrast, is typically conducted by the builders themselves, which cannot deliver that impartial scrutiny.

An AI Risk Dashboard aggregates governance metrics into executive-level views, surfacing overall risk posture and emerging trends across AI systems. This satisfies the stem's requirement for leadership oversight: directors need consolidated, strategic visibility rather than operational detail, enabling informed decisions on risk appetite, resourcing and regulatory exposure without inspecting individual model logs.

Inference request latency measures the elapsed time between submitting input and receiving the model's prediction. Monitoring it exposes performance-related AI risk such as throughput degradation or infrastructure saturation, satisfying the stem's requirement for a standard performance monitoring metric.

A model card documents intended use, training data, evaluation metrics, limitations and ethical considerations. Auditors rely on it as the authoritative record of how the AI system was designed, validated and governed, satisfying the evidence requirement for an AI audit.

A common risk taxonomy supplies shared categories and definitions, so AI risks are identified, communicated and reported using identical terminology across business units. This directly satisfies the stem's requirement for consistent enterprise-wide risk management rather than fragmented, unit-specific registers.

Mapping AI risk metrics onto existing enterprise risk categories lets AI appetite be expressed in the same taxonomy ERM already aggregates and reports. This satisfies the alignment requirement by integrating AI exposure into established operational, financial, and reputational reporting rather than running a parallel framework.

Anomaly scores from embedding or activation layers detect OOD inputs because they measure deviation in the model's learned feature representation, not just output confidence. This satisfies the stem's requirement for early lifecycle risk detection, since distributional shift appears in internal representations before predictions degrade.

Third-party outsourcing transfers execution but not accountability. Contractual audit rights, model validation requirements and risk-sharing clauses preserve the organisation's ability to inspect vendor models, verify performance claims and allocate liability, keeping governance effective across the vendor boundary.

ISO 42001 emphasizes stakeholder communication and consultation.

Incomplete datasets with many missing values directly degrade training data quality, since gaps bias model learning and reduce statistical representativeness. This satisfies the stem's training-phase constraint by addressing data completeness, one of the three recognised contributors to Data Quality Risk alongside accuracy and consistency issues during model development.

An AI Bill of Materials documents training data lineage, model provenance and dependencies, giving visibility into inherited third-party risks. This control satisfies the supply chain constraint by enabling traceability and vulnerability assessment of externally sourced models, which contractual clauses alone cannot provide.

Highly complex models, such as deep networks with many parameters, behave as opaque systems, so tracing why a specific output occurred becomes impractical. That opacity directly obstructs interpretability and root cause analysis when incidents arise during operation.

Data preparation and training is where datasets are assembled, split and fed to models. If temporal ordering or target variables leak across train and test boundaries here, future or label information contaminates training, making this stage the primary leakage point.

Drift incident frequency and severity translate technical model degradation into business risk terms boards govern: how often performance decays and how much harm results. This quantifies operational stability and oversight effectiveness better than granular statistical metrics.

An AI Governance/Ethics Committee provides the cross-functional coordination the stem demands, uniting legal, data science, compliance and business stakeholders to triage bias incidents. Bias issues span model, data and policy domains, so routing first to this committee satisfies the requirement for coordinated oversight rather than single-team remediation.

Black-box AI models resist conventional ISO 31000 assessment because their reasoning is opaque. Adding an explainability and interpretability assessment step lets assessors evaluate how outputs are produced, satisfying the framework's need to identify and analyse AI-specific risks.

Fairness constraints embedded in the training objective directly penalise biased outcomes during optimisation, so the model learns equitable decision boundaries rather than having bias patched afterwards. This satisfies the mitigation requirement by reducing disparate impact at its source, before deployment.

When transparency cannot be obtained, risk management requires explicit recognition of the residual risk plus compensating controls. Contractual attestations create accountability, independent testing rights permit future validation, and challenger models with outcome monitoring provide empirical assurance. This approach satisfies audit expectations by making the limitation visible to governance bodies and bounding its impact, rather than pretending the risk does not exist.

Centralised AI governance routes all approvals, reviews and policy decisions through one body, so requests queue behind limited reviewers. That concentration creates operational bottlenecks and delays project timelines, which is the principal trade-off versus decentralised, domain-embedded governance.

Documenting the test methodology, success criteria and mitigation steps proves the bias test was rigorous and that identified disparities were remediated, not merely measured. This satisfies the governance policy's audit requirement by evidencing both the assessment approach and the corrective action taken.

Third-party library vulnerabilities in an AI model fall under the AI Risk Manager and Security team, who own risk identification, assessment and mitigation across the AI supply chain. This satisfies the stem's accountability question by assigning ownership to those governing AI risk rather than developers alone.