Isaca AAIA Practice Test 45
Question 440
An independent AI auditor reviews this case: A product recommendation model is corrupted by a burst of fake reviews introduced through one third-party feed. Which TWO complementary elements should the auditor seek: an effective corrective control and the most probative independent evidence? Select TWO answers.
- Compare ingestion source identities, sampling patterns, hash manifests, anomaly screens and approval records.
- Train directly on all externally submitted records because the model is expected to learn robustly by itself.
- Validate provenance and anomaly patterns for incoming data, quarantine suspicious updates and approve retraining sources.
- Separate development, test and production roles with controlled promotion, scoped credentials and audit trails.
- Inspect generation method, similarity and disclosure tests, subgroup distributions and downstream use-case validation.
Explanation
Correct Answers
A:The reason it is necessary is: Training-data integrity is a control boundary because malicious or flawed inputs can change model behavior persistently. For this situation: A product recommendation model is corrupted by a burst of fake reviews introduced through one third-party feed.
C:The reason it is necessary is: Training-data integrity is a control boundary because malicious or flawed inputs can change model behavior persistently. This response fits the risk associated with Data poisoning prevention in automated ingestion: Validate provenance and anomaly patterns for incoming data, quarantine suspicious updates and approve retraining sources.
Incorrect Answers
B:It misses the key control principle: Training-data integrity is a control boundary because malicious or flawed inputs can change model behavior persistently.
D:This option concerns AI data and model CI/CD separation, which is a related but different AI governance concern. The decisive principle here is: Training-data integrity is a control boundary because malicious or flawed inputs can change model behavior persistently.
E:This option concerns Synthetic training data utility and privacy controls, which is a related but different AI governance concern. The decisive principle here is: Training-data integrity is a control boundary because malicious or flawed inputs can change model behavior persistently.
Question 441
A governance and assurance review uncovers the following: A public chatbot uses submitted support articles as training data without source authenticity checks. What is the MOST appropriate risk-based audit recommendation that addresses the root control weakness? Select ONE best answer.
- Train directly on all externally submitted records because the model is expected to learn robustly by itself.
- Enforce point-in-time joins and temporal cutoffs for training and evaluation features.
- Validate provenance and anomaly patterns for incoming data, quarantine suspicious updates and approve retraining sources.
- Validate synthetic data for utility, subgroup representation, memorization and plausible disclosure before release.
Explanation
Correct Answer
C:The reason it is necessary is: Training-data integrity is a control boundary because malicious or flawed inputs can change model behavior persistently. In this case: A public chatbot uses submitted support articles as training data without source authenticity checks.
Incorrect Answers
A:This attractive shortcut does not resolve the Data poisoning prevention in automated ingestion weakness: Train directly on all externally submitted records because the model is expected to learn robustly by itself. In this case: A public chatbot uses submitted support articles as training data without source authenticity checks.
B:This option concerns Feature-store time leakage in model evaluation, which is a related but different AI governance concern. In this case: A public chatbot uses submitted support articles as training data without source authenticity checks.
D:This option concerns Synthetic training data utility and privacy controls, which is a related but different AI governance concern. In this case: A public chatbot uses submitted support articles as training data without source authenticity checks.
Question 442
A risk-based audit engagement examines this operating condition: A public chatbot uses submitted support articles as training data without source authenticity checks. Which principle BEST explains why the observed practice is a material governance or control concern? Select ONE best answer.
- Synthetic labels do not automatically eliminate privacy, bias or representativeness risk.
- Accuracy on imbalanced data can obscure missed rare high-consequence cases or overwhelming false positives.
- Train directly on all externally submitted records because the model is expected to learn robustly by itself.
- Training-data integrity is a control boundary because malicious or flawed inputs can change model behavior persistently.
Explanation
Correct Answer
D:Independent audit evidence includes: Compare ingestion source identities, sampling patterns, hash manifests, anomaly screens and approval records. In this case: A public chatbot uses submitted support articles as training data without source authenticity checks.
Incorrect Answers
A:This option concerns Synthetic training data utility and privacy controls, which is a related but different AI governance concern. For this situation: A public chatbot uses submitted support articles as training data without source authenticity checks.
B:This option concerns Class imbalance and minority event coverage, which is a related but different AI governance concern. In this case: A public chatbot uses submitted support articles as training data without source authenticity checks.
C:This attractive shortcut does not resolve the Data poisoning prevention in automated ingestion weakness: Train directly on all externally submitted records because the model is expected to learn robustly by itself. For this situation: A public chatbot uses submitted support articles as training data without source authenticity checks.
Question 443
An audit committee asks for a defensible assessment of this situation: A public chatbot uses submitted support articles as training data without source authenticity checks. What evidence should the auditor examine FIRST to substantiate or challenge the control owner’s claim? Select ONE best answer.
- Inspect index ACL propagation, chunk metadata, identity claims, denied-search tests and cached answer behavior.
- Compare ingestion source identities, sampling patterns, hash manifests, anomaly screens and approval records.
- Inspect generation method, similarity and disclosure tests, subgroup distributions and downstream use-case validation.
- Train directly on all externally submitted records because the model is expected to learn robustly by itself.
Explanation
Correct Answer
B:This evidence directly tests the asserted control for Data poisoning prevention in automated ingestion: Compare ingestion source identities, sampling patterns, hash manifests, anomaly screens and approval records.
Incorrect Answers
A:This option concerns RAG corpus access controls and tenant segregation, which is a related but different AI governance concern. The decisive principle here is: Training-data integrity is a control boundary because malicious or flawed inputs can change model behavior persistently.
C:This option concerns Synthetic training data utility and privacy controls, which is a related but different AI governance concern. The audit must corroborate: Compare ingestion source identities, sampling patterns, hash manifests, anomaly screens and approval records.
D:For this situation: A public chatbot uses submitted support articles as training data without source authenticity checks.
Question 444
During an AI controls assessment, the team learns that a public chatbot uses submitted support articles as training data without source authenticity checks. After management addresses the issue, which result MOST convincingly demonstrates effective remediation? Select ONE best answer.
- Authorized queries work while cross-tenant and restricted-content retrievals are blocked and logged.
- Threshold and sample design meet documented detection and false-intervention goals for rare events.
- Train directly on all externally submitted records because the model is expected to learn robustly by itself.
- Contaminated feeds are contained and verified clean data can be restored without corrupting promoted models.
Explanation
Correct Answer
D:This is an observable validation result for Data poisoning prevention in automated ingestion, not just management assurance: Contaminated feeds are contained and verified clean data can be restored without corrupting promoted models.
Incorrect Answers
A:This option concerns RAG corpus access controls and tenant segregation, which is a related but different AI governance concern. For this situation: A public chatbot uses submitted support articles as training data without source authenticity checks.
B:This option concerns Class imbalance and minority event coverage, which is a related but different AI governance concern. For this situation: A public chatbot uses submitted support articles as training data without source authenticity checks.
C:The auditor should instead examine: Compare ingestion source identities, sampling patterns, hash manifests, anomaly screens and approval records.
Question 445
An independent AI auditor reviews this case: A default model trains with delinquency features calculated a month after its labeled prediction date. What is the MOST appropriate risk-based audit recommendation that addresses the root control weakness? Select ONE best answer.
- Enforce point-in-time joins and temporal cutoffs for training and evaluation features.
- Evaluate class distribution with relevant precision, recall and business costs, and use justified stratification or targeted samples.
- Validate provenance and anomaly patterns for incoming data, quarantine suspicious updates and approve retraining sources.
- Accept impressive validation results without checking whether features encode outcomes from the future.
Explanation
Correct Answer
A:The reason it is necessary is: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference. The acceptable verified outcome is: Held-out evaluation uses only information available at the decision time and predicts realistic production performance.
Incorrect Answers
B:This option concerns Class imbalance and minority event coverage, which is a related but different AI governance concern. It does not establish control effectiveness for Feature-store time leakage in model evaluation.
C:This option concerns Data poisoning prevention in automated ingestion, which is a related but different AI governance concern. It does not establish control effectiveness for Feature-store time leakage in model evaluation.
D:This attractive shortcut does not resolve the Feature-store time leakage in model evaluation weakness: Accept impressive validation results without checking whether features encode outcomes from the future.
Question 446
A governance and assurance review uncovers the following: A default model trains with delinquency features calculated a month after its labeled prediction date. Which principle BEST explains why the observed practice is a material governance or control concern? Select ONE best answer.
- Accept impressive validation results without checking whether features encode outcomes from the future.
- Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference.
- A model can disclose information statistically through outputs even without returning underlying records verbatim.
- Accuracy on imbalanced data can obscure missed rare high-consequence cases or overwhelming false positives.
Explanation
Correct Answer
B:The acceptable verified outcome is: Held-out evaluation uses only information available at the decision time and predicts realistic production performance. This is the applicable AI audit principle for Feature-store time leakage in model evaluation: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference.
Incorrect Answers
A:This attractive shortcut does not resolve the Feature-store time leakage in model evaluation weakness: Accept impressive validation results without checking whether features encode outcomes from the future. For this situation: A default model trains with delinquency features calculated a month after its labeled prediction date.
C:This option concerns Model inversion and sensitive attribute leakage, which is a related but different AI governance concern. It does not establish control effectiveness for Feature-store time leakage in model evaluation.
D:This option concerns Class imbalance and minority event coverage, which is a related but different AI governance concern. The decisive principle here is: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference.
Question 447
A risk-based audit engagement examines this operating condition: A default model trains with delinquency features calculated a month after its labeled prediction date. What evidence should the auditor examine FIRST to substantiate or challenge the control owner’s claim? Select ONE best answer.
- Accept impressive validation results without checking whether features encode outcomes from the future.
- Inspect feature timestamps, label windows, join keys, leakage tests and reconstructed historical decisions.
- Inspect index ACL propagation, chunk metadata, identity claims, denied-search tests and cached answer behavior.
- Inspect output schemas, reconstruction probes, privacy threat model, aggregation and access restrictions.
Explanation
Correct Answer
B:The reason it is necessary is: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference. This evidence directly tests the asserted control for Feature-store time leakage in model evaluation: Inspect feature timestamps, label windows, join keys, leakage tests and reconstructed historical decisions.
Incorrect Answers
A:This attractive shortcut does not resolve the Feature-store time leakage in model evaluation weakness: Accept impressive validation results without checking whether features encode outcomes from the future. The auditor should instead examine: Inspect feature timestamps, label windows, join keys, leakage tests and reconstructed historical decisions.
C:This option concerns RAG corpus access controls and tenant segregation, which is a related but different AI governance concern. It does not establish control effectiveness for Feature-store time leakage in model evaluation.
D:This option concerns Model inversion and sensitive attribute leakage, which is a related but different AI governance concern. The decisive principle here is: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference.
Question 448
An audit committee asks for a defensible assessment of this situation: A default model trains with delinquency features calculated a month after its labeled prediction date. After management addresses the issue, which result MOST convincingly demonstrates effective remediation? Select ONE best answer.
- Accept impressive validation results without checking whether features encode outcomes from the future.
- Labels are consistent against approved definitions, and model tests expose residual uncertain classes.
- Contaminated feeds are contained and verified clean data can be restored without corrupting promoted models.
- Held-out evaluation uses only information available at the decision time and predicts realistic production performance.
Explanation
Correct Answer
D:The reason it is necessary is: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference. This is an observable validation result for Feature-store time leakage in model evaluation, not just management assurance: Held-out evaluation uses only information available at the decision time and predicts realistic production performance.
Incorrect Answers
A:This attractive shortcut does not resolve the Feature-store time leakage in model evaluation weakness: Accept impressive validation results without checking whether features encode outcomes from the future. Effective closure requires: Held-out evaluation uses only information available at the decision time and predicts realistic production performance.
B:This option concerns Label quality and human annotator agreement, which is a related but different AI governance concern. It does not establish control effectiveness for Feature-store time leakage in model evaluation.
C:This option concerns Data poisoning prevention in automated ingestion, which is a related but different AI governance concern. The decisive principle here is: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference.
Question 449
During an AI controls assessment, the team learns that a medical risk forecast uses post-discharge records that would be unknown when clinicians originally made decisions. What is the MOST appropriate risk-based audit recommendation that addresses the root control weakness? Select ONE best answer.
- Validate provenance and anomaly patterns for incoming data, quarantine suspicious updates and approve retraining sources.
- Define annotation rubrics, representative blind review, disagreement resolution and documented quality thresholds.
- Enforce point-in-time joins and temporal cutoffs for training and evaluation features.
- Accept impressive validation results without checking whether features encode outcomes from the future.
Explanation
Correct Answer
C:The reason it is necessary is: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference. In this case: A medical risk forecast uses post-discharge records that would be unknown when clinicians originally made decisions.
Incorrect Answers
A:This option concerns Data poisoning prevention in automated ingestion, which is a related but different AI governance concern. In this case: A medical risk forecast uses post-discharge records that would be unknown when clinicians originally made decisions.
B:This option concerns Label quality and human annotator agreement, which is a related but different AI governance concern. The decisive principle here is: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference.
D:This attractive shortcut does not resolve the Feature-store time leakage in model evaluation weakness: Accept impressive validation results without checking whether features encode outcomes from the future. In this case: A medical risk forecast uses post-discharge records that would be unknown when clinicians originally made decisions.
Question 450
An independent AI auditor reviews this case: A medical risk forecast uses post-discharge records that would be unknown when clinicians originally made decisions. Which TWO complementary elements should the auditor seek: an effective corrective control and the most probative independent evidence? Select TWO answers.
- Inspect generation method, similarity and disclosure tests, subgroup distributions and downstream use-case validation.
- Evaluate class distribution with relevant precision, recall and business costs, and use justified stratification or targeted samples.
- Accept impressive validation results without checking whether features encode outcomes from the future.
- Enforce point-in-time joins and temporal cutoffs for training and evaluation features.
- Inspect feature timestamps, label windows, join keys, leakage tests and reconstructed historical decisions.
Explanation
Correct Answers
D:This response fits the risk associated with Feature-store time leakage in model evaluation: Enforce point-in-time joins and temporal cutoffs for training and evaluation features.
E:The reason it is necessary is: Future information leakage creates unrealistic validation metrics and cannot be tolerated in prospective inference. The corrective action is: Enforce point-in-time joins and temporal cutoffs for training and evaluation features.
Incorrect Answers
A:This option concerns Synthetic training data utility and privacy controls, which is a related but different AI governance concern. It does not establish control effectiveness for Feature-store time leakage in model evaluation.
B:This option concerns Class imbalance and minority event coverage, which is a related but different AI governance concern. In this case: A medical risk forecast uses post-discharge records that would be unknown when clinicians originally made decisions.
C:This attractive shortcut does not resolve the Feature-store time leakage in model evaluation weakness: Accept impressive validation results without checking whether features encode outcomes from the future. For this situation: A medical risk forecast uses post-discharge records that would be unknown when clinicians originally made decisions.
