About these questions
These five questions are original. We wrote them in the style of the AAISM content outline, with one best answer out of four, which is the format ISACA describes in its candidate guide. They are not real exam questions, and the real exam will word things differently. All five are single-answer questions.
How to use these
Try each question without notes, and choose an answer before you open the explanation. Then read the explanation even when you were right, because the reason the closest wrong option fails is where the learning is. Note the domain where you slipped and revisit it on the domains explained page.
Question 1 (AI Governance and Program Management)
A hospital ethics committee asks for a document that describes how the images used to train a triage model were collected, which patient groups appear in them, and where the data should not be reused. The committee is not asking how well the trained model performs. Which artifact should the security manager provide?
- A. A model card for the trained model
- B. A datasheet describing the dataset
- C. An AI bill of materials
- D. An incident report for the training pipeline
Show the answer and explanation
Correct answer: B. A dataset datasheet documents how data was gathered and composed, who is represented in it, and what its known limits and unsuitable uses are. That is exactly what the committee requested. A model card, the closest wrong option, describes the trained model: intended use, performance, and limitations, which the committee explicitly said it does not want. An AI bill of materials lists components rather than describing a dataset’s composition. An incident report covers an event that has happened, and none is mentioned.
Question 2 (AI Risk Management)
A support chatbot answers questions using documents from a shared folder to which any employee can upload files. Attackers are known to hide instructions inside uploaded documents. If one succeeds, customer refund records could be disclosed. Which element of this description is the vulnerability?
- A. The people who hide instructions in documents
- B. The disclosure of customer refund records
- C. The folder accepting uploads from any employee
- D. The chatbot answering support questions
Show the answer and explanation
Correct answer: C. A vulnerability is a weakness that a threat can use. Here, the weakness is that content the chatbot trusts can be changed by anyone. The attackers are the threat, since they are the actor with the potential to cause harm, and the disclosure of refund records is the impact. The chatbot answering questions is the intended function, not a weakness in itself. Getting the three parts right matters because the treatment differs: you control the source of the content, not the existence of the attackers.
Question 3 (AI Risk Management)
A bank has applied all reasonable controls to its credit-scoring model. One low-likelihood, very costly failure mode remains. The finance director asks whether buying cover for the resulting financial loss can be part of the treatment plan. Which answer is MOST appropriate?
- A. No, because insurance cannot be part of risk treatment
- B. Yes, because insurance removes the failure mode from the register
- C. Yes, as a transfer of financial impact that does not lower the likelihood
- D. No, unless the model’s vendor has bought the same cover
Show the answer and explanation
Correct answer: C. Transfer is a recognized treatment option, and insurance suits exposure that stays material after mitigation, particularly when the event is unlikely but expensive. It funds the loss; it does not make the failure less likely, and it does not remove accountability or fix reputational harm. Option B is the closest wrong answer because it overstates what insurance does. Option A is wrong because transfer is a legitimate treatment. Option D adds a condition that has nothing to do with whether the bank may treat its own residual exposure this way.
Question 4 (AI Technologies and Controls)
A legal assistant searches one shared index of case files. After it drafts an answer, a filter removes quoted passages the lawyer is not cleared to see. Reviewers find that the summaries still reveal restricted facts. Which design change is MOST appropriate?
- A. Enforce the asking user’s permissions at retrieval time
- B. Make the filter on the final answer stricter
- C. Tell the model in its instructions to leave out restricted material
- D. Retrieve fewer passages for each question
Show the answer and explanation
Correct answer: A. If a restricted passage reaches the model, the model can paraphrase or summarize it, and a filter that only checks quoted text will miss that. Authorization must be applied when documents are retrieved, using the rights of the person asking, so that unauthorized material never enters the context. A stricter output filter, the closest wrong option, is still a late and unreliable control. Instructing the model relies on a probabilistic system to cooperate. Retrieving fewer passages reduces exposure a little but does not enforce any permission.
Question 5 (AI Technologies and Controls)
One engineer at a retailer prepares the training data, trains the model, signs off on the evaluation results, and publishes the model to the registry that production pulls from. No one else reviews any step. Which control is missing?
- A. Rate limiting on the registry
- B. Encryption of the training data at rest
- C. Longer retention of training logs
- D. Separation of duties between building, approving, and deploying
Show the answer and explanation
Correct answer: D. Separation of duties divides the roles so that no single person can move a model into production unchecked, whether by mistake or on purpose. In a machine learning pipeline, that means independent evaluation, a distinct approval role, and controlled promotion into the registry. Encryption at rest, the closest wrong option, protects confidentiality, but the problem here is the lack of independent review. Rate limiting addresses load, and longer log retention helps investigations after the fact; neither stops the single point of failure.
More guides for AAISM
- AAISM Exam Format, Scoring and Retakes
- AAISM Study Plan: 4-Week and 8-Week
- Is the AAISM Exam Hard?
- AAISM Prerequisites: CISM or CISSP First
- AAISM Renewal and Retakes Explained
- Is AAISM Worth It? An Honest Look
- AAISM Cheat Sheet: Key Facts to Review
- AAIA vs AAISM: Which ISACA AI Exam First?
- AAIR vs AAISM: Which ISACA AI Exam First?
Next step
If you missed a question, go to the matching section of domains explained. For more practice at exam length, our AAISM practice tests have 450 original questions across all three domains.