Isaca AAISM Online Practice Questions and Exam Preparation

Isaca AAISM Online Questions & Answers

• Question 181:

  Which of the following is a key risk indicator (KRI) for an AI system used for threat detection?

  A. Number of training epochs
  B. Training time of the model
  C. Number of layers in the neural network
  D. Number of system overrides by cyber analysts
  D. Number of system overrides by cyber analysts

  ---

  AAISM materials emphasize that in operational AI systems, key risk indicators (KRIs) must reflect risks to performance and reliability rather than technical design factors alone. In the case of threat detection, the most relevant KRI is the frequency of system overrides by human analysts, as this indicates a lack of trust, frequent false positives, or poor detection accuracy. Training epochs, model depth, and training time are technical metrics but do not directly measure operational risk. Analyst overrides

  represent a practical measure of system effectiveness and risk.

  References:

  AAISM Study Guide - AI Risk Management (Operational KRIs for AI Systems)

  ISACA AI Security Management - Monitoring AI Effectiveness

• Question 182:

  An organization is reviewing an AI application to determine whether it is still needed. Engineers have been asked to analyze the number of incorrect predictions against the total number of predictions made. Which of the following is this an example of?

  A. Control self-assessment (CSA)
  B. Model validation
  C. Key performance indicator (KPI)
  D. Explainable decision-making
  C. Key performance indicator (KPI)

  ---

  AAISM guidance identifies metrics like error rate versus total predictions as a key performance indicator (KPI) for evaluating AI model effectiveness. KPIs provide measurable values to assess performance against objectives. Model validation is broader and occurs prior to production use, testing the model against predefined standards. Control self-assessment relates to governance processes, not predictive accuracy.

  Explainable decision-making refers to interpretability, not error-rate evaluation. Thus, analyzing incorrect

  predictions against total predictions is a performance measure, making it a KPI.

  References:

  AAISM Exam Content Outline - AI Governance and Program Management (Performance Metrics and KPIs)

  AI Security Management Study Guide - Accuracy and Error Metrics

• Question 183:

  A financial organization is concerned about AI data poisoning. Which control BEST mitigates this risk?

  A. Implementing a break-glass policy
  B. Transparency with customers about data sources
  C. Using training data from multiple sources
  D. Delivering AI-specific security awareness training
  C. Using training data from multiple sources

  ---

  AAISM outlines that diversifying training data sources reduces the likelihood and impact of poisoning because: attack samples become harder to inject anomalies are easier to detect corrupted data has reduced influence on model behavior Break-glass procedures (A) relate to incident response, not mitigation. Customer transparency (B) does not stop poisoning. Awareness training (D) is insufficient alone.

  References: AAISM Study Guide - Data Poisoning Mitigation; Data Diversity and Source Validation.

• Question 184:

  The PRIMARY benefit of implementing moderation controls in generative AI applications is that it can:

  A. Increase the model's ability to generate diverse and creative content
  B. Optimize the model's response time
  C. Ensure the generated content adheres to privacy regulations
  D. Filter out harmful or inappropriate content
  D. Filter out harmful or inappropriate content

  ---

  AAISM materials identify the primary benefit of moderation controls in generative AI systems as their ability to filter out harmful, offensive, or inappropriate content before it is delivered to users. This safeguards organizational reputation, compliance, and user trust. While moderation may indirectly support compliance with privacy requirements, its main function is ensuring that outputs align with ethical and safety standards.

  Moderation does not enhance creativity or response speed. Its primary value is in controlling the quality of

  generated outputs by blocking harmful content.

  References:

  AAISM Study Guide - AI Technologies and Controls (Moderation and Output Controls)

  ISACA AI Security Management - Harmful Content Mitigation in Generative AI

• Question 185:

  A data scientist creating categories and training the algorithm on large data sets is an example of which type of AI model learning technique?

  A. Reinforcement
  B. Unsupervised
  C. Machine learning (ML)
  D. Supervised
  D. Supervised

  ---

  AAISM classifies learning paradigms by the presence of labeled targets. Creating categories (labels) and training on them is supervised learning, where input features are mapped to known outputs and optimization minimizes prediction error against ground truth. Unsupervised (B) discovers structure without labels; reinforcement (A) optimizes behavior via rewards; "machine learning" (C) is the broad field, not the specific technique.

  References: AI Security Management (AAISM) Body of Knowledge - AI/ML Foundations; Learning Paradigms and Data Requirements. AAISM Study Guide - Supervised vs. Unsupervised vs. Reinforcement Learning; Label Quality and Model Performance Dependencies.

• Question 186:

  Which of the following strategies BEST ensures generative AI tools do not expose company data?

  A. Conducting an independent AI data audit
  B. Testing AI tools before implementation
  C. Implementing a solution to prohibit the input of sensitive data
  D. Ensuring AI tools are compliant with local regulations
  C. Implementing a solution to prohibit the input of sensitive data

  ---

  AAISM prioritizes preventive controls at the point of use for generative AI, specifically input-governance and DLP controls that block or redact confidential, regulated, or high-risk data before it can be sent to external models. Audits, pre-deployment tests, and regulatory conformance are necessary but do not themselves prevent an employee from pasting sensitive content into prompts. Enforcing input restrictions, pattern-based redaction, policy-aware controls, and allow-lists for approved contexts provides the

  highest assurance of preventing exposure.

  References:. AI Security Management. (AAISM) Body of Knowledge: Data loss prevention for AI; prompt /input controls; approved channels and guardrails for generative AI.. AI Security Management. Study Guide: Preventive over detective controls for confidentiality; enterprise guardrails at prompt capture and egress points.

• Question 187:

  The PRIMARY goal of data poisoning attacks is to:

  A. compromise the confidentiality of output data from the model
  B. compromise the confidentiality of model input data
  C. manipulate the behavior of the model during development
  D. undermine the integrity of the AI system's outputs
  D. undermine the integrity of the AI system's outputs

  ---

  AAISM defines data poisoning as the insertion of malicious or corrupted data into training (or fine-tuning) pipelines to degrade or bias model behavior, thereby compromising output integrity in production. While poisoning occurs during development/training (C), its primary objective is the downstream integrity impact on predictions/outputs (D). Options A and B relate to confidentiality threats (e.g., inversion or leakage), not poisoning.

  References:

  AI Security Management (AAISM) Body of Knowledge: Model Integrity Threats-data poisoning aims and effects; supply-side data controls

  AAISM Study Guide: Risk scenarios for poisoning; mitigations via data provenance checks, robust training, and anomaly detection

• Question 188:

  Within an incident handling process, which of the following would BEST help restore end-user trust in an AI system?

  A. Remediation of the AI system based on lessons learned
  B. The AI model's outputs are validated by team members
  C. AI is used to monitor incident detection and alerts
  D. The AI model prioritizes incidents based on business impact
  A. Remediation of the AI system based on lessons learned

  ---

  AAISM highlights that post-incident remediation and demonstrating lessons learned is essential to restoring trust. Governance guidance specifies that stakeholders regain confidence only when organizations show clear corrective actions, transparency, and improvements to prevent recurrence. Validating outputs (B) supports accuracy but is not trust-restoring. Monitoring (C) and prioritization (D) relate to operations, not trust rebuilding.

  References: AAISM Study Guide - AI Governance; Incident Response and Trust Restoration.

• Question 189:

  Which of the following controls BEST mitigates the risk of data poisoning?

  A. Data set restoration
  B. Data validation
  C. Digital watermarking
  D. Intrusion detection
  B. Data validation

  ---

  The AAISM technical controls framework emphasizes data validation as the primary safeguard against data poisoning attacks. Poisoning occurs when attackers insert malicious or corrupted data into training sets. Validation techniques verify the quality, authenticity, and consistency of input data before training, preventing compromised samples from corrupting the model. Restoration helps after compromise, watermarking protects ownership, and intrusion detection monitors networks rather than data quality.

  The most effective preventive measure is data validation.

  References:

  AAISM Study Guide - AI Technologies and Controls (Data Poisoning Mitigation)

  ISACA AI Security Management - Data Validation and Quality Controls

• Question 190:

  During adversarial red-team testing, which technique BEST simulates an evasion attack?

  A. Injecting mislabeled data into the training pipeline
  B. Modifying input samples to bypass detection
  C. Corrupting the model weights
  D. Performing model extraction via APIs
  B. Modifying input samples to bypass detection

  ---

  AAISM defines evasion attacks as inference-time manipulations of inputs to avoid detection (e.g., tweaking malware payloads to appear benign). Mislabeled training data (A) is poisoning. Weight corruption (C) is sabotage. Model extraction (D) is IP theft.

  References:

  AAISM Body of Knowledge ?Evasion Attacksand; Input Manipulation Techniques.