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

Stopping the risk is the priority, followed by a root cause analysis to adjust the governance or training processes.

Infrastructure risks focus on availability and security of the compute environment.

The time required to complete a risk assessment for a new model tracks how streamlined the governance process is.

Shadow AI refers to AI tools deployed without the knowledge or oversight of the governance team, creating invisible, unmanaged risks.

Statistical distribution shifts (like KL Divergence) in input data serve as a proxy for performance degradation.

Non-deterministic training data results in unstable models where performance can fluctuate unpredictably.

Controls must cover the lifecycle, from development integrity to ongoing monitoring and access management.

Documentation and training parameters are crucial for reproducing and auditing model behavior.

Transfer learning risks include the inheritance of latent biases from the pre-trained weights.

The Data Protection Officer (DPO) and Legal/Compliance teams are responsible for regulatory compliance.

Legal ensures compliance with evolving regulations, while Data Science/Engineering provides the technical understanding of how the model functions.

An inventory is the foundational requirement for knowing what models exist, their criticality, and their risk profile.

Risk appetite is a strategic decision involving both business leaders (who own the risk) and risk managers (who define the framework).

The AI Risk Appetite Statement is the document approved by the Board that defines the risk tolerance for AI activities.

If no one is designated as accountable for the model's outcomes, the governance structure is non-existent or dysfunctional.

Risk assessment involves identifying risks and then evaluating them based on impact/likelihood.

A Data Protection Impact Assessment (DPIA) or an AI-specific Risk Assessment protocol is the formal process for identifying risks early.

Data lineage, schema validation, and outlier detection are standard data engineering controls.

Once a threshold is breached, the model should be taken offline or placed under remediation until it is back within acceptable risk parameters.

The cycle consists of identification, assessment/treatment, and ongoing monitoring.

Data splitting prevents overfitting and provides an unbiased estimate of the model's performance on unseen data.

The Chief Risk Officer (CRO) or equivalent enterprise risk lead ensures AI risk is treated as a component of overall enterprise risk.

NIST AI RMF 'Govern' function involves establishing the culture and processes for AI risk management, including clear definition of roles and accountability. Documenting the designated 'Model Owner' and their specific risk accountabilities is a core component.

Automated validation of performance metrics against a 'golden dataset' ensures the model meets the required threshold.

Directors must understand AI risks and benefits to provide informed oversight and define the organization's risk appetite.

An AI inventory is the starting point for any risk program, as you cannot manage risks for models you haven't identified.

Risk appetite statements provide the boundary of acceptable risk, which is foundational to any AI risk management program.

Red Teaming is an advanced 'Measure' or 'Verification' activity designed to find weaknesses, bias, or safety risks in a model.

Concept drift indicates that the model is no longer operating in the environment it was trained for, leading to inaccurate predictions and potential risk.

The Chief Risk Officer (CRO) ensures that AI development aligns with the risk appetite and risk management framework, acting as the bridge between technical execution and business oversight.

Risk mitigation is most effective during the Design phase, before the model is developed, to avoid costly re-engineering.

The AI Governance Policy or Charter establishes the roles, responsibilities, and decision-making authorities for AI risk.

Guardrails are technical controls placed on top of models to intercept and filter output, which is a vital part of risk governance.

Adoption requires alignment with existing processes and clear communication.

Before integrating, the organization must understand the AI scope and criticality relative to existing assets.

Audit trails and model lineage are critical for meeting regulatory transparency and accountability standards.

Diverse data sourcing and fairness evaluation are standard methods for bias mitigation.

Performance metrics do not account for external risks such as legal exposure, compliance violations, or social impact.

Inconsistent formatting and lack of temporal relevance are clear data quality issues.

Standardizing risk language, establishing clear reporting lines, and aligning risk criteria are essential for ERM integration.

Aligning risk appetite requires mapping technical AI performance metrics to business-level risk registers, ensuring that performance trade-offs are explicitly approved by stakeholders.

The Ethics Committee provides advisory, values-based guidance, while the Governance body maintains decision-making authority for risk.

Executives need to see how AI risk affects business outcomes (e.g., financial impact, reputation) rather than just technical performance.

Model lineage/provenance trackers record the exact data versions and training parameters used for a specific model version.

Board oversight requires metrics on accountability, such as the auditability of model decision-making trails.

Risk management must be integrated at the beginning (Design/Requirement) to avoid costly redesigns later.

Microsoft Purview's AI hub requires activity logging to be enabled to audit interactions with generative AI applications.

The NIST AI RMF 'Govern' function prioritizes understanding the provenance and lineage to establish accountability and transparency.

Board reports need to focus on strategic risk alignment and the efficacy of the oversight process, specifically tracking the risk management pipeline and significant residual risks.

Establishing clear boundaries for where generative AI is acceptable versus prohibited is a key aspect of risk appetite.

AI regulations, such as the EU AI Act, differ significantly by jurisdiction, making local legal expertise essential for compliance risk management.

The 'Govern' function focuses on the culture, policies, and structures that foster a risk-aware AI lifecycle.

Validation is the independent verification that a model meets its intended purpose and risk performance requirements.

Dashboards allow leadership to see aggregated risk trends across the enterprise, not just individual project metrics.

Latency and error rates are the primary operational indicators of model performance risk.

A model card provides the necessary metadata, performance limitations, and intended use cases required for auditability.

A common taxonomy ensures that AI risk is understood and communicated consistently across the enterprise.

Integrating AI risk appetite into the existing ERM framework ensures consistency and avoids silos.

OOD detection identifies inputs that are fundamentally different from the distribution the model was trained on.

Third-party risk management must include audit rights and clear accountability for the vendor's models.

ISO 42001 requires systematic risk identification, analysis, and treatment planning.

Incomplete data, selection bias, and incorrect labeling all directly degrade the model's foundation.

Vendor/Model transparency (Software Bill of Materials for AI) is critical for understanding dependencies and hidden risks in black-box models.

Extremely complex models (e.g., deep neural networks) are often 'black boxes', making them hard to interpret and verify.

Data leakage typically occurs during the data preparation and feature engineering stages of the AI lifecycle.

The frequency and severity of model drift incidents are direct indicators of operational and reputational risk, which are board-level concerns.

The AI Governance team or AI Ethics office is tasked with immediate triage of AI-specific risks before escalation to legal or IT operations.

Inclusion of model explainability assessments is necessary to address the opacity inherent in deep learning models.

Mitigation involves both pre-processing (data) and in-processing (training) techniques.

Centralization can create 'bottlenecks' that slow down innovation and delay deployment.

Documenting the bias test methodology and results is essential to demonstrate due diligence to regulators.

The AI Risk Manager, working with the security team, must assess the risk and ensure the library is updated or the dependency is mitigated.

Risk should be linked to strategic objectives and the organization's appetite to ensure the Board understands the impact.

Version control is essential for reproducibility and rollback, minimizing the risk of unrecoverable failures.

Accountability ensures that individuals or teams are responsible for the outcomes of AI systems, preventing a 'blame-free' environment for AI failures.