This is not a replacement for the official study guide. Please read the official study guide first — use this exam prep only to reinforce or practice your understanding.
Free AAISM exam practice questions with answers and explanations, organized by domain and part.
Domain 1: AI Governance and Program Management
Q311 An LLM is being attacked with abusive prompts. Which containment technique is MOST appropriate?
- A. Delete the audit logs to prevent attackers from reading earlier prompts
- B. Retrain the underlying model from scratch before any further prompts are accepted
- C. Disconnect all employees from email until the abuse has stopped entirely
- D. Deploy input and output validation and prompt templates to screen abuse prompts
Answer: D
Input and output validation and prompt templates screen and sanitize abusive prompts and shield the model from direct attacks. Retraining may be impractical and costly. Deleting logs destroys evidence.
Domain 1: AI Governance and Program Management
Q312 A data poisoning attack is in progress against a training pipeline. What should the team do FIRST to contain it?
- A. Publish the dataset externally so that others can identify tampered records
- B. Revoke access to datasets and preprocessing scripts for systems in the pipeline
- C. Increase the training epochs so the model overwrites the poisoned patterns
- D. Disable audit logging on the pipeline to reduce noise during the investigation
Answer: B
Containing poisoning means revoking access to datasets for systems in the pipeline and to preprocessing code to stop further tampering. More training or publishing the data worsens the situation.
Domain 1: AI Governance and Program Management
Q313 Which control is MOST effective to contain systematic adversarial inference attacks via an API?
- A. Increasing the model size to absorb the probing queries
- B. Adding more training data to make the model harder to infer
- C. Hiding the training dataset location from all API consumers
- D. Throttling, validating and sanitizing data inputs
Answer: D
These attacks use validated input channels such as APIs, so throttling, validation and sanitization reduce them. Larger models or more data do not stop systematic probing.
Domain 1: AI Governance and Program Management
Q314 Why is eradication of an AI incident MOST difficult compared with patching a conventional vulnerability?
- A. It depends on the attack technique and may require retraining or removing poisoned data
- B. Changing credentials always suffices because attackers only target access keys and never the model or its data
- C. Vendor patches always fix model behavior without any retraining or data review
- D. AI incidents never need eradication because models reset themselves on restart
Answer: A
Conventional incidents can be eradicated by patching or credential changes, whereas AI eradication depends on the technique, such as retraining or removing poisoned data and rebuilding. Retraining for LLMs may be impractical.
Domain 1: AI Governance and Program Management
Q315 After removing poisoned data from a training set, what is the MOST appropriate next step?
- A. Increase the learning rate so the model adapts to new inputs faster
- B. Delete all evidence of the poisoned data to prevent future access
- C. Retrain or rebuild the model with the cleaned data
- D. Resume production with the existing model since the data is now clean
Answer: C
Eradication of poisoning requires isolating and removing the poisoned data, then training a new model on clean data. Resuming unchanged leaves the poisoned model in place. Evidence should be preserved.
Domain 1: AI Governance and Program Management
Q316 A security operations center is considering AI-powered incident response. Which challenge deserves MOST attention?
- A. Models never need updating once they have been trained on historical incidents
- B. AI always eliminates the need for human analysts in the response process
- C. Bias cannot occur in security models because attack data is purely technical
- D. False positives and inaccuracies can reduce analyst trust
Answer: D
AI can produce inaccurate results or false positives, which erode trust, so continuous monitoring and refinement are required. Human oversight remains critical. Biased or stale models are also risks.
Domain 1: AI Governance and Program Management
Q317 Why is human oversight MOST important in AI-powered incident response?
- A. Because regulation prohibits any automation of incident response activities
- B. To ensure automated actions align with business objectives and ethical standards
- C. Because AI cannot process security data without a human typing each query
- D. To slow every response deliberately so that analysts can verify every alert
Answer: B
Oversight balances automation with human input to avoid unintended consequences and keep actions aligned with objectives and ethics. AI can process data, and automation is not generally prohibited.
Domain 1: AI Governance and Program Management
Q318 A SOC's AI model is not updated with recent threat intelligence. What is the MOST likely consequence?
- A. The model becomes ineffective against emerging attack vectors
- B. Alerts become more accurate because the model keeps stable assumptions
- C. Analysts need less training since the tool behaves predictably over time
- D. Attackers stop evolving once defenses use consistent detection patterns
Answer: A
Failing to keep models current with threat intelligence can make them ineffective against emerging risk. Static models do not improve accuracy. Attackers continue evolving.
Domain 1: AI Governance and Program Management
Q319 Training data for a security AI model omits several attack types. Which risk is MOST likely?
- A. Perfect detection of new attacks because the model has no prior assumptions
- B. Lower alert volume with no downside because fewer attack types are modeled
- C. Overlooked threats or disproportionate responses due to biased or incomplete data
- D. Faster eradication of all incidents because the model focuses on known cases
Answer: C
Models are only as good as their training data; biased or incomplete data skews results, causing overlooked threats or disproportionate responses. It does not create perfect detection.
Domain 2: AI Risk Management
Q320 An AI project team finds it hard to say when and why the system might fail. Which approach to risk does the manual recommend MOST for such systems?
- A. A purely technical review of source code before every release of the model, with sign-off from the engineering manager alone
- B. A risk-based approach that considers societal and trustworthiness factors, not just technical ones
- C. A single annual vulnerability scan across all deployed models
- D. An exemption from formal risk assessment because AI adapts on its own
Answer: B
AI systems learn and evolve unpredictably, so risk reaches beyond technical issues and a risk-based approach is needed. Code review or scanning cannot capture societal factors. Adaptive systems need more, not less, risk assessment.