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 3: AI Technologies and Controls
Q571 A team is designing the hosting for a new AI platform and must decide how to handle security from the beginning. Which approach is MOST consistent with secure by design?
- A. Treat security as a final review step just before go-live
- B. Embed security requirements, threat modeling and logging into the architecture before building
- C. Rely on the cloud provider to add controls without any customer input
- D. Build the platform first and add controls after the first release, which regulators have generally tolerated so far
Answer: B
Secure by design builds protections in from inception through end of life. Adding controls later or delegating them entirely leaves gaps that are costly to close.
Domain 3: AI Technologies and Controls
Q572 An architect stores embeddings of confidential documents in a vector database. Which concern is MOST important to address?
- A. Access to the embeddings and metadata must respect the sensitivity of the source documents
- B. Embeddings are random numbers so they carry no sensitivity at all, with results reported at the next quarterly meeting
- C. Only the original files need protection since embeddings cannot be reversed
- D. Vector databases cannot be secured with identity-based access controls
Answer: A
Embeddings and their metadata can reveal sensitive content, so access controls must reflect the source classification. Assuming they are harmless ignores inversion and retrieval risk.
Domain 3: AI Technologies and Controls
Q573 A company wants to prevent an AI service from being exposed to the whole internet. Which architectural control is the BEST choice?
- A. Open inbound ports so partners can test without delay
- B. A public endpoint protected only by an obscure URL, and later audits can confirm it if needed and in line with how the team normally works
- C. Network segmentation with private endpoints and gateway authentication in front of the service
- D. A shared API key printed in the documentation
Answer: C
Segmentation, private endpoints and authenticated gateways limit exposure. An obscure URL is not a control, and open ports or shared keys increase the attack surface.
Domain 3: AI Technologies and Controls
Q574 A model is retrained weekly and each version behaves slightly differently. Which practice would MOST help an auditor reconstruct what ran on a given date?
- A. Informal notes kept by the developer who last trained the model
- B. Screenshots taken when the team remembers to capture them
- C. A single folder where the newest model replaces the previous one, because it is the quickest path to the stated goal
- D. Versioned records linking each deployed model to its data, code and configuration
Answer: D
Versioned lineage lets the auditor trace each deployment to its inputs. Overwriting, informal notes and ad hoc screenshots cannot reliably reconstruct past behavior.
Domain 3: AI Technologies and Controls
Q575 A business unit wants to swap in a newer foundation model to reduce cost. What should happen FIRST before the change goes live?
- A. A change request with testing, risk review and approval under the AI change process
- B. A switch followed by a review only if users complain, which keeps the schedule predictable for sponsors
- C. An immediate switch since the new model is advertised as better
- D. Approval by the vendor's sales team on the unit's behalf
Answer: A
Replacing a model can change behavior, risk and compliance, so it needs testing and authorization through change management. Advertising claims and after-the-fact review skip validation.
Domain 3: AI Technologies and Controls
Q576 During planning, a team has defined goals but has not yet considered who could be harmed by the system. Which action is the BEST next step?
- A. Skip the assessment because the use case is internal only
- B. Start training so real results can reveal the risks later
- C. Ask the vendor to assess the harm on the team's behalf, which regulators have generally tolerated so far
- D. Run an initial risk and ethical impact assessment with stakeholders before building
Answer: D
Planning and design should include early risk and ethical impact assessment with stakeholders. Building first or delegating assessment to the vendor leaves the organization unaware of its own risk.
Domain 3: AI Technologies and Controls
Q577 A data team plans to train on customer emails collected years ago. Which question should they ask FIRST?
- A. Do the original consent and notices permit use of this data for AI training?
- B. Is the dataset large enough to improve the leaderboard score?
- C. Can the emails be formatted as spreadsheets?, with results reported at the next quarterly meeting
- D. Are the emails stored in the fastest available storage tier?
Answer: A
Purpose and consent determine whether the data can lawfully be used for training. Size, storage speed and format are engineering concerns that come after the legal basis is confirmed.
Domain 3: AI Technologies and Controls
Q578 Testing shows that a model's error rate is much higher for one customer segment. What is the MOST appropriate action before release?
- A. Investigate the data and model for bias, fix or mitigate it, and document the residual risk
- B. Add a disclaimer to the marketing page and proceed
- C. Remove the segment from the test set so the metric improves, which keeps the schedule predictable for sponsors
- D. Release the model because the overall average accuracy is acceptable
Answer: A
A subgroup disparity indicates potential bias that should be investigated and mitigated, with documented risk acceptance. Averages hide harm, and removing test data or adding disclaimers does not fix it.
Domain 3: AI Technologies and Controls
Q579 A pilot of an AI tool is about to expand to all staff. Which step is the MOST important before full rollout?
- A. Let each department decide separately whether to adopt the tool
- B. Review pilot results against success criteria and authorize the rollout through governance
- C. Expand immediately since the pilot period has ended, with results reported at the next quarterly meeting
- D. Skip training to avoid delaying the launch
Answer: B
Pilot findings should be assessed against defined criteria and approved through governance. Automatic expansion and uncoordinated adoption bypass oversight, and skipping training raises misuse risk.
Domain 3: AI Technologies and Controls
Q580 A model in production starts to produce outputs that its owners did not intend. Which mechanism is MOST useful in the operate and monitor phase?
- A. Annual reviews that only look at uptime, as long as the budget for the work stays unchanged
- B. Feedback channels and monitoring that trigger action by the solution owner
- C. A policy that discourages users from reporting issues
- D. Daily restarts of the model server
Answer: B
Established feedback structures and monitoring that trigger owner action detect unintended impacts. Uptime reviews and discouraged reporting hide problems, and restarts do not change behavior.