View Full Isaca AAIR Exam Dumps and Practice Test Dumps.

Q241. An enterprise plans several AI initiatives, but management is uncertain which projects deserve funding. Which consideration should have the GREATEST influence on prioritization from an AI risk perspective?

1. The number of developers requesting the project
   2. Whether the AI technology is currently popular
   3. Expected business value considered together with risk exposure and strategic alignment
   4. The size of the underlying model

Correct Answer: 3. Expected business value considered together with risk exposure and strategic alignment

Explanation: AI initiatives should be evaluated in the context of value creation as well as risk. A technically impressive project may not justify investment if it solves a low-priority business problem or creates exposure outside organizational risk appetite. Management should assess expected benefits, strategic relevance, affected stakeholders, legal and ethical implications, operational requirements, and residual risk. This enables resources to be directed toward AI use cases that provide meaningful organizational value within acceptable boundaries. ISACA specifically emphasizes evaluating AI use cases according to organizational objectives and risk appetite rather than pursuing AI technology simply because it is innovative.

Q242. An organization wants teams to experiment with AI without exposing production data or customers to uncontrolled risk. What is the BEST governance approach?

1. Establish a controlled AI sandbox with defined data, access, testing, and promotion requirements
   2. Permit unrestricted experimentation directly in production
   3. Prohibit all AI experimentation
   4. Allow each employee to select their own public AI platform

Correct Answer: 1. Establish a controlled AI sandbox with defined data, access, testing, and promotion requirements

Explanation: A controlled sandbox supports innovation while limiting exposure. It can restrict sensitive data, external connectivity, user access, autonomous actions, and integration with production systems while allowing teams to test ideas. Governance should also define when an experimental solution must undergo formal risk assessment, validation, security review, and business approval before promotion. This approach balances value creation with risk management. Unrestricted production experimentation can expose customers and regulated data, while banning experimentation entirely may unnecessarily constrain innovation. A sandbox creates an environment where experimentation occurs within explicitly defined boundaries.

Q243. A risk professional discovers that two regulatory requirements appear to impose different documentation expectations for the same high-risk AI system. What should be done FIRST?

1. Apply whichever requirement is easier
   2. Ignore both until a regulator asks for evidence
   3. Allow the development team to choose
   4. Analyze applicability and reconcile the requirements with legal and compliance stakeholders**

Correct Answer: 4. Analyze applicability and reconcile the requirements with legal and compliance stakeholders

Explanation: Regulatory obligations can overlap, differ in scope, or appear inconsistent. The organization should first determine which requirements apply based on jurisdiction, industry, data, affected parties, and AI use case. Legal and compliance specialists can then determine how the obligations should be reconciled and whether one common control or documentation approach can satisfy both. Selecting the easier requirement without analysis can create noncompliance. The resulting interpretation should be documented and incorporated into the organization’s regulatory inventory, control mapping, and AI lifecycle processes so project teams can apply it consistently.

Q244. A generative AI system provides career recommendations to employees. Which governance issue is MOST important if employees believe the system may disadvantage particular groups?

1. Whether the model uses the newest architecture
   2. Fairness, transparency, and mechanisms for human review or challenge
   3. The physical location of the development servers
   4. The number of tokens produced per response

Correct Answer: 2. Fairness, transparency, and mechanisms for human review or challenge

Explanation: AI systems influencing careers, opportunities, or other meaningful outcomes should be assessed for fairness and potential discriminatory effects. Employees should also understand the role of AI in the recommendation process and have an appropriate means to raise concerns or obtain human review. Transparency and redress are especially important when outputs may affect advancement or development opportunities. Model architecture and response length do not determine whether the process is trustworthy. Governance should consider stakeholder impact, legal obligations, bias testing, oversight, and accountability throughout the system’s lifecycle.

Q245. A predictive model uses a variable that would not actually be available at the time a real-world decision is made, but the variable is included during training. What risk does this MOST directly create?

1. Vendor lock-in
   2. Model inversion
   3. Target or information leakage that can inflate apparent model performance
   4. Disaster recovery failure

Correct Answer: 3. Target or information leakage that can inflate apparent model performance

Explanation: Information leakage occurs when model development uses information that would not legitimately be available at prediction time. The model may then achieve unrealistically strong validation results because it indirectly receives clues about the outcome it is supposed to predict. Once deployed, that information is missing and real-world performance can collapse. Development and validation processes should examine feature timing, data lineage, preprocessing, and the exact information available at inference. Strong headline accuracy is not reliable evidence when the training design contains leakage. Preventing leakage is therefore an important part of model validation and data governance.

Q246. An organization wants to reproduce the exact results of an earlier model-training run. Which information is MOST important to retain?

1. Only the final accuracy score
   2. Only the model owner’s name
   3. Only the final model file
   4. Versioned code, training data, parameters, dependencies, environment, and relevant randomization settings**

Correct Answer: 4. Versioned code, training data, parameters, dependencies, environment, and relevant randomization settings

Explanation: Reproducibility requires sufficient information to reconstruct how a model was created. The final model artifact alone does not show which data version, preprocessing logic, libraries, hyperparameters, random seeds, or execution environment produced it. Retaining these elements supports investigations, independent validation, model comparison, rollback, and regulatory evidence. The level of detail should reflect system criticality and applicable retention requirements. Reproducibility is particularly important when a historical decision is challenged or when management needs to determine why two apparently similar training runs produced different results.

Q247. An AI system has passed all planned tests, but reviewers discover that the test environment does not represent actual production data volumes. What should happen before deployment?

1. Perform additional validation using production-representative scale and operating conditions
   2. Deploy because functional testing already passed
   3. Increase risk tolerance to accommodate uncertainty
   4. Remove performance requirements

Correct Answer: 1. Perform additional validation using production-representative scale and operating conditions

Explanation: Validation should reflect the conditions under which an AI system will actually operate. A model or supporting application can function correctly at small scale yet fail when data volume, concurrency, latency, or infrastructure demand increases. Production-representative testing can reveal capacity limits, timeouts, degraded accuracy, or unexpected interactions. The depth of testing should be proportionate to criticality and expected scale. Passing a nonrepresentative test does not provide sufficient evidence that a high-impact production deployment will remain reliable under realistic operating conditions.

Q248. A model update is ready for production. Which condition would provide the STRONGEST basis for an automatic rollback?

1. A developer dislikes the new user interface
   2. A predefined material post-deployment performance or risk threshold is breached
   3. The model version number is higher
   4. More users access the system than yesterday

Correct Answer: 2. A predefined material post-deployment performance or risk threshold is breached

Explanation: Rollback decisions should be based on objective, predefined conditions rather than ad hoc preference. Relevant triggers may include accuracy deterioration, fairness violations, unacceptable error rates, safety failures, latency thresholds, or other material risk indicators. Predefining these conditions makes response faster and reduces debate during a production incident. Rollback procedures should also be tested and ensure that the previous version, dependencies, and data remain compatible. This provides a corrective lifecycle control when post-deployment monitoring identifies behavior that differs materially from approved expectations.

Q249. An attacker sends extremely long prompts designed to consume excessive compute and degrade availability of a generative AI service. What threat does this MOST closely represent?

1. Model or resource denial-of-service attack
   2. Data minimization
   3. Model versioning
   4. Risk acceptance

Correct Answer: 4. Model or resource denial-of-service attack

Explanation: Generative AI systems can be vulnerable to resource-exhaustion attacks in which adversaries submit computationally expensive prompts or high request volumes to consume inference resources and degrade service. Controls may include request-size limits, quotas, authentication, rate limits, workload isolation, anomaly detection, resource monitoring, and graceful degradation. This differs from model extraction, which seeks to reproduce model behavior, or prompt injection, which attempts to manipulate instructions. Availability threats should be incorporated into AI threat assessments, incident response, and capacity planning when services are externally accessible or business critical.

Q250. A risk workshop has identified a potential AI failure but has not documented how the failure could cause business harm. What is missing from the scenario?

1. A model version number
   2. The consequence or business impact pathway
   3. A vendor invoice
   4. The development language

Correct Answer: 1. The consequence or business impact pathway

Explanation: A useful risk scenario should explain not only what can go wrong but also why it matters to the organization. The scenario should connect a threat, failure, vulnerability, or adverse condition to affected assets or processes and then to plausible business consequences. Those consequences can include financial loss, harm to individuals, regulatory exposure, service interruption, or reputational damage. Without an impact pathway, management cannot reliably assess severity or choose proportional treatment. AI risk should therefore be expressed in business terms rather than as a list of isolated technical weaknesses.

Q251. An AI risk exists before any safeguards have been considered. What does this represent?

1. Residual risk
   2. Transferred risk
   3. Inherent risk
   4. Accepted risk only

Correct Answer: 3. Inherent risk

Explanation: Inherent risk represents the level of exposure that exists before considering the effect of controls or other treatment activities. Understanding inherent risk helps management identify which scenarios have the greatest raw potential impact and how much reliance is being placed on controls. Residual risk is the exposure remaining after those controls and treatments are considered. Comparing inherent and residual risk can also support assessment of control effectiveness. AI risk registers should distinguish these concepts consistently so reporting accurately reflects both the underlying exposure and the contribution of implemented safeguards.

Q252. A risk treatment plan relies on a compensating control until a preferred technical control can be implemented. What is MOST important?

1. Assume the compensating control is automatically equivalent
   2. Remove the original control requirement permanently
   3. Stop measuring residual risk
   4. Validate that the compensating control reduces the relevant risk sufficiently and define a review period**

Correct Answer: 4. Validate that the compensating control reduces the relevant risk sufficiently and define a review period

Explanation: A compensating control should address the same underlying risk to an acceptable degree rather than merely provide the appearance of compliance. Its design and operating effectiveness should be assessed, and residual risk should be recalculated. Because compensating controls are often temporary, the organization should define ownership, monitoring, and a review or expiration date. If the alternative does not reduce risk within tolerance, additional treatment or escalation is required. Temporary measures should not silently become permanent simply because implementation of the preferred control is delayed.

Q253. Which evidence provides the STRONGEST support that an AI access control is operating consistently over time?

1. A policy stating access must be restricted
   2. A developer’s written assurance
   3. Periodic operating evidence and testing showing unauthorized access is prevented or detected
   4. A model accuracy report

Correct Answer: 2. Periodic operating evidence and testing showing unauthorized access is prevented or detected

Explanation: Control documentation demonstrates intended design, but operating effectiveness requires evidence that the control works consistently. Relevant evidence might include access logs, denied requests, periodic access reviews, technical configuration, exception records, and independent test results. The appropriate evidence depends on whether the control is preventive, detective, or corrective. A developer’s statement is weaker than objective evidence, and model accuracy does not demonstrate access-control effectiveness. Material AI controls should be monitored and tested with frequency proportionate to risk and change.

Q254. A control fails repeatedly because staff do not understand when it should be used. Which corrective action is MOST appropriate?

1. Clarify the procedure, provide role-specific training, and reassess whether the control design is practical
   2. Increase the risk score without investigating
   3. Remove all training requirements
   4. Ignore the failures because the control exists

Correct Answer: 1. Clarify the procedure, provide role-specific training, and reassess whether the control design is practical

Explanation: Repeated operating failures can result from unclear procedures, insufficient training, poor workflow integration, or impractical control design. Corrective action should address the root cause rather than merely documenting that failure occurred. The organization may need to simplify the process, automate parts of the control, clarify responsibilities, or provide focused training. After remediation, control effectiveness should be retested. A control that exists only on paper does not adequately reduce AI risk when employees cannot consistently apply it in real operating conditions.

Q255. Management wants a metric that warns when unresolved high-risk AI control deficiencies are accumulating. Which indicator is MOST appropriate?

1. Number of AI marketing campaigns
   2. Percentage or count of overdue high-risk control deficiencies
   3. Number of AI developers
   4. Average model file size

Correct Answer: 2. Percentage or count of overdue high-risk control deficiencies

Explanation: Overdue high-risk deficiencies directly indicate that material control weaknesses remain unresolved beyond agreed timelines. Monitoring their number, age, severity, and responsible owners provides useful information about risk-treatment effectiveness. Rising backlogs can signal inadequate resources, poor accountability, or systemic implementation problems. The indicator should be linked to escalation requirements so severe or long-overdue weaknesses receive appropriate management attention. Metrics unrelated to control remediation provide little insight into whether known AI risk exposures are being addressed.

Q256. An enterprise wants management reports to distinguish between current risk exposure and future emerging AI threats. What is the BEST approach?

1. Report only historical incidents
   2. Combine current KRIs with forward-looking horizon-scanning and emerging-risk information
   3. Remove current risk indicators
   4. Treat every emerging threat as an existing incident

Correct Answer: 3. Combine current KRIs with forward-looking horizon-scanning and emerging-risk information

Explanation: Current risk indicators provide evidence about existing exposure, while horizon scanning helps management anticipate changes that may not yet appear in operational metrics. Combining both provides a more complete risk picture. Emerging-risk information can include new attack techniques, regulations, vendor developments, technological changes, and industry incidents. These should not automatically be treated as realized incidents, but they can trigger scenario development and reassessment. AI evolves rapidly, making forward-looking risk intelligence especially important alongside conventional monitoring.

Q257. A vendor supplies a critical AI API but does not provide guaranteed performance during periods of peak demand. What contractual risk should be addressed?

1. Capacity and performance service-level expectations during material demand conditions
   2. The vendor’s office decoration policy
   3. Employee training schedules
   4. The customer’s model-development language

Correct Answer: 4. Capacity and performance service-level expectations during material demand conditions

Explanation: Availability alone may not be sufficient when business processes depend on timely AI responses. A technically available service can still fail operational requirements if latency or throughput degrades severely during peak demand. Contracts should define relevant performance expectations, measurement methods, escalation, and remedies based on service criticality. The organization should also test its own capacity assumptions and maintain appropriate fallback procedures. Performance risk belongs within third-party and continuity planning when degraded service can materially affect customer or business outcomes.

Q258. A third-party AI provider has excellent controls but is the organization’s only feasible supplier for a critical capability. What should management recognize?

1. Strong vendor controls eliminate all third-party risk
   2. Concentration and substitutability risk may remain even when the vendor itself is well controlled
   3. Exit planning is unnecessary
   4. The supplier can automatically accept the organization’s risk

Correct Answer: 2. Concentration and substitutability risk may remain even when the vendor itself is well controlled

Explanation: A strong supplier can still create material enterprise concentration risk when no practical substitute exists. Outages, regulatory restrictions, geopolitical events, acquisitions, price increases, or strategic changes may leave the organization with few alternatives. Risk assessment should therefore distinguish the vendor’s control quality from the organization’s dependency on the vendor. Management may need stronger continuity planning, contractual protections, data portability, alternate architectures, or formal risk acceptance even when supplier due diligence results are positive.

Q259. During an AI incident, responders discover that an automated agent executed actions that were never included in the approved use case. What should occur during the post-incident review?

1. Focus only on restoring service
   2. Delete the agent’s activity logs
   3. Assess authorization boundaries, tool permissions, use-case controls, and why scope restrictions failed
   4. Increase agent privileges to avoid future errors

Correct Answer: 1. Assess authorization boundaries, tool permissions, use-case controls, and why scope restrictions failed

Explanation: If an AI agent acts outside its approved scope, the incident review should determine how that was technically possible. Investigation should examine tool permissions, workload identity, policy enforcement, prompt handling, approval mechanisms, and whether intended-use restrictions were implemented only through documentation rather than enforceable controls. Lessons learned should feed into authorization design, testing, monitoring, and governance. Restoring service is necessary but does not address the underlying control failure that allowed excessive agency.

Q260. A disaster recovery exercise restores a critical AI service within the required RTO, but several hours of recent transaction data are missing beyond the approved RPO. How should the test be assessed?

1. Fully successful because the service restarted on time
   2. Successful because model accuracy is unaffected
   3. Irrelevant because RPO does not apply to AI services
   4. Partially unsuccessful because the recovery point objective was not met**

Correct Answer: 4. Partially unsuccessful because the recovery point objective was not met

Explanation: RTO and RPO measure different recovery requirements. Meeting the RTO means the service was restored within the required timeframe, while missing more data than allowed means the RPO was not achieved. Both objectives should be evaluated independently. The organization should investigate backup frequency, replication, data synchronization, or recovery sequencing and improve the plan. An AI service may rely on transaction records, feature data, logs, model state, or other information whose loss can affect accuracy, auditability, or business processing after recovery.