Free AAISM sample questions with answers
These follow ISACA’s house style: scenario-led, from the accountable manager’s perspective, with a qualifier that decides the answer. Read the capitalised word carefully.
Question 1. An enterprise has discovered that several business units are using public generative AI tools with company data, without approval. Which should the security manager do FIRST?
- A. Issue a policy prohibiting all public generative AI tools
- B. Deliver awareness training to all staff
- C. Identify and inventory the AI tools in use and the data involved
- D. Block access to all public AI services at the network perimeter
Show answer
Answer: C
You cannot govern what you have not identified, so establishing the inventory and the extent of use comes before policy, training or blocking. A policy written without knowing what is in use will not fit, blocking prematurely drives usage further underground, and training addresses behaviour before you know its scope.
Question 2. A vendor supplies a pretrained model that will be fine-tuned on customer data and deployed in a regulated process. Which is the MOST important element of the vendor assessment?
- A. The vendor’s pricing model and licensing terms
- B. Provenance and integrity of the model and its training data
- C. The quality of the vendor’s API documentation
- D. How frequently the vendor releases new model versions
Show answer
Answer: B
Provenance and integrity of the model and its training data determine whether the resulting system can be trusted or explained to a regulator, and they cannot be established after deployment. Pricing, interface quality and release frequency are legitimate procurement concerns but none of them addresses whether the model itself is trustworthy.
Question 3. An AI system supporting credit decisions has begun producing materially different outcomes from six months ago, with no code change. Which is the BEST explanation to investigate first?
- A. Model drift relative to the original training data
- B. A network latency problem between services
- C. An expired TLS certificate on the endpoint
- D. A change in the vendor’s licence agreement
Show answer
Answer: A
Model drift, where performance shifts as real-world data diverges from training data, explains changed behaviour without a code change and is the first thing to examine. A network issue would affect availability rather than decision quality, an expired certificate would break connectivity, and a licence change does not alter outputs.
Question 4. The board asks for assurance that a new AI system will not cause harm to customers before it goes live. Which should the security manager recommend?
- A. A penetration test of the hosting environment
- B. A business continuity plan for the AI service
- C. Post-release monitoring with monthly reporting
- D. A pre-deployment impact assessment with defined human oversight
Show answer
Answer: D
An impact assessment before deployment, with defined human oversight of consequential decisions, is the control that produces the assurance the board is asking for and is the recognised pre-deployment governance artefact. A penetration test addresses technical exploitability, a continuity plan addresses outage, and post-release monitoring provides evidence only after the harm could occur.
Question 5. Which control MOST directly reduces the risk that a model reproduces confidential information from its training data in responses to users?
- A. Encryption of the training data at rest
- B. Data minimisation in the training set combined with output filtering
- C. Multifactor authentication for all users of the system
- D. A longer log retention period for all model interactions
Show answer
Answer: B
Output filtering and data minimisation in the training set act on the two points where the disclosure actually occurs: what went in, and what comes out. Encryption at rest protects stored data rather than generated responses, MFA controls who can access the system rather than what it reveals, and a longer retention period increases exposure.
How did you do?
Notice how few of these have a technical answer. Even question 5, from the technologies domain, is about which control addresses a stated risk rather than how to configure anything. That is ISACA’s style throughout — you are the accountable manager, and the answer is usually a decision, an assessment or a sequence.
If you found yourself picking defensible-but-not-best options, that is the most common failure mode for this exam and it is trainable. Work through the study plan, paying attention to the section on question style.