AAISM
Can you pass this 10-question AAISM quiz?
Ten questions taken from our AAISM Practice Questions, three or four from each of ISACA's three domains, with four options each, as on the real exam. Tap an answer to see why, with the source. Give yourself the exam's pace: 100 seconds a question.
Can you get 10 out of 10?
1 of 10Over 6 months, your company has approved five AI pilots. Each time, business units argued about who could approve a new use case, who could accept its risk and who could stop a deployment. Which action would BEST stop the same argument from returning with the next model?
Show answer
Answer: A. The AI governance charter is the approved document that sets up AI oversight. It is where you settle, once, who approves new AI use cases, who accepts AI risk and who can stop a deployment. Without it, each question is argued again for every new model. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), Domain 1, AI Governance and Program Management; task 1 (charter, roles and responsibilities)
2 of 10Your AI awareness course shows a 98% completion rate. In the same month, a developer pasted a customer database export into a public coding assistant. Which measure would BEST show whether the training is working?
Show answer
Answer: C. Completion and quiz scores show that people clicked through, not that they changed what they do. The signals that show whether training worked come from your controls. Look at DLP events on AI destinations, shadow AI found by discovery and incidents reported by staff rather than found by tools. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), 1D AI Security Program Development and Management; tasks 18 and 21 (security metrics, awareness training)
3 of 10Early on a Saturday you learn that the sales agent, which is allowed to send follow-up emails, has spent the night sending customers discount codes that were never approved. The emails are still going out. Hidden text on a website the agent read is the suspected cause. What should you do FIRST?
Show answer
Answer: A. When harm is ongoing, containment comes before root cause analysis. Choose the smallest action that stops the harm: removing the abused capability ends the emails and keeps the agent answering questions. If you could not tell which capability was being abused, you would switch the whole system off. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), 1E Business Continuity and Incident Response; task 16 (containment, escalation, eradication and recovery)
4 of 10The data science lead reports that a claims-triage model is 97% accurate and concludes that the likelihood of a harmful error is low. The model triages 20,000 claims a month. What should you do NEXT?
Show answer
Answer: A. Accuracy is a performance measure on a test set, not a likelihood. At 20,000 claims a month, a 3% error rate means about 600 wrong decisions a month. What matters is how many errors fall in a harmful direction, how each segment fares and whether controls catch the errors before they land. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), 2A AI Risk Assessment, Thresholds, and Treatment
5 of 10Human reviewers in a claims process can change or reject the model's recommendations. Over 3 months the override rate has fallen close to zero. What should you do NEXT?
Show answer
Answer: A. The override rate needs a floor as well as a ceiling. A rate near zero may mean the model is excellent, or that the reviewers have stopped looking, and only a sample check of approved cases tells you which. Automation bias can turn a reviewer from a control into a formality. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), 2A AI Risk Assessment, Thresholds, and Treatment; task 6 (factors that call for reassessment of risk)
6 of 10A vendor launched an AI drafting assistant in its product last quarter. As evidence of control, it sends you a SOC 2 Type 2 report on its platform whose period ended before the assistant launched. Which statement about this report is accurate?
Show answer
Answer: C. Every report has a boundary, set by its scope statement and its period. A report whose period ended before the vendor launched the assistant does not cover the assistant, however sound the platform controls were. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), 2C AI Vendor and Supply Chain Management; task 9 (verify AI security requirements in vendor AI-enabled solutions)
7 of 10A scheduling agent got caught in a loop overnight. It sent 400 meeting invitations and made 9,000 model calls before anyone noticed. Which control would have bounded this incident?
Show answer
Answer: C. Hard limits in code hold whatever the model decides. Cap the number of steps an agent may take, the money it may spend and the rate at which it may act, and give it a kill switch. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), 3A AI Security Architecture and Design
8 of 10An architect proposes letting the model decide whether a user may open a record, because the user's role is written in the prompt. Under a zero trust design, which component should decide and enforce that access?
Show answer
Answer: A. Zero trust separates the component that decides whether a request is allowed from the component that enforces the decision. In an AI system the model is neither: its context mixes your instructions with content strangers wrote, so it cannot be trusted to decide or to enforce. Your identity and policy services decide, and the gateway, the orchestration code and the downstream systems enforce. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), 3A AI Security Architecture and Design; task 10 (security architecture specifically for AI)
9 of 10A policy assistant answers staff questions from retrieved passages. When the passages do not cover a question, it still gives a fluent, confident answer that is wrong. Which design change BEST reduces this?
Show answer
Answer: A. Grounding constrains an answer to verifiable source content and checks that the content supports it. Abstention lets the system decline, or say it does not know, when support falls below a threshold. Together they act on the cause, which is an answer with nothing behind it. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), 3C Data Management Controls
10 of 10The alert count from your AI gateway falls by 80% in one week. The gateway team has reported no change. What should you do FIRST?
Show answer
Answer: B. A sudden drop in alerts can mean the attacks stopped. It can also mean the gateway stopped seeing traffic, or that someone changed or switched off its rules. Monitor the monitor: check the event stream for gaps, look for configuration changes and confirm that the tool is still receiving what it should. Source: ISACA, AAISM Exam Candidate Guide v1.26, Examination Content Outline (June 2025), 3E Security Controls and Monitoring
AAISM flashcards for Anki
front,back "Over 6 months, your company has approved five AI pilots. Each time, business units argued about who could approve a new use case, who could accept its risk and who could stop a deployment. Which action would BEST stop the same argument from returning with the next model? | A) Agree an AI governance charter that settles approval, risk acceptance and stop authority once. / B) Build a RACI matrix for each pilot as it arrives at the governance committee. / C) Ask the security team to approve every new use case and every deployment until the disputes subside. / D) Add a clause to the AI acceptable use policy naming security as the approver.","A) Agree an AI governance charter that settles approval, risk acceptance and stop authority once.. The AI governance charter is the approved document that sets up AI oversight. It is where you settle, once, who approves new AI use cases, who accepts AI risk and who can stop a deployment. Without it, each question is argued again for every new model. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz" "Your AI awareness course shows a 98% completion rate. In the same month, a developer pasted a customer database export into a public coding assistant. Which measure would BEST show whether the training is working? | A) The share of staff who passed the course quiz on their first attempt. / B) The number of AI training modules completed per employee this year. / C) The trend in DLP events on AI destinations and in AI incidents reported by staff. / D) The share of new hires who finished the course within their first 30 days.","C) The trend in DLP events on AI destinations and in AI incidents reported by staff.. Completion and quiz scores show that people clicked through, not that they changed what they do. The signals that show whether training worked come from your controls. Look at DLP events on AI destinations, shadow AI found by discovery and incidents reported by staff rather than found by tools. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz" "Early on a Saturday you learn that the sales agent, which is allowed to send follow-up emails, has spent the night sending customers discount codes that were never approved. The emails are still going out. Hidden text on a website the agent read is the suspected cause. What should you do FIRST? | A) Remove the agent's permission to send email and leave its other functions running. / B) Find the website content that carried the hidden instruction and confirm the root cause. / C) Notify the affected customers and the regulator that unapproved codes were sent. / D) Roll the agent back to its previous model version and restart it under watch.","A) Remove the agent's permission to send email and leave its other functions running.. When harm is ongoing, containment comes before root cause analysis. Choose the smallest action that stops the harm: removing the abused capability ends the emails and keeps the agent answering questions. If you could not tell which capability was being abused, you would switch the whole system off. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz" "The data science lead reports that a claims-triage model is 97% accurate and concludes that the likelihood of a harmful error is low. The model triages 20,000 claims a month. What should you do NEXT? | A) Ask for error volumes, the direction of errors, results by segment and tested controls. / B) Record the likelihood as low and cite the accuracy figure as the evidence. / C) Reject the model, since a 3% error rate is too high for claims work. / D) Calculate an annualized loss expectancy from the 3% error rate.","A) Ask for error volumes, the direction of errors, results by segment and tested controls.. Accuracy is a performance measure on a test set, not a likelihood. At 20,000 claims a month, a 3% error rate means about 600 wrong decisions a month. What matters is how many errors fall in a harmful direction, how each segment fares and whether controls catch the errors before they land. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz" "Human reviewers in a claims process can change or reject the model's recommendations. Over 3 months the override rate has fallen close to zero. What should you do NEXT? | A) Sample-check a set of cases the reviewers approved. / B) Remove the review step, since the model's output plainly no longer needs correction. / C) Treat the rate as a KPI success and cut how often override figures are reported. / D) Report the rate to the governance committee as evidence that the model performs well.","A) Sample-check a set of cases the reviewers approved.. The override rate needs a floor as well as a ceiling. A rate near zero may mean the model is excellent, or that the reviewers have stopped looking, and only a sample check of approved cases tells you which. Automation bias can turn a reviewer from a control into a formality. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz" "A vendor launched an AI drafting assistant in its product last quarter. As evidence of control, it sends you a SOC 2 Type 2 report on its platform whose period ended before the assistant launched. Which statement about this report is accurate? | A) A Type 2 report covers the design of controls at a single point in time, not over a period / B) A SOC 2 report is the vendor's own attestation rather than an independent assurance report / C) The report's period ended before the assistant existed, so it does not cover the assistant / D) Because it is a Type 2 report, its coverage extends to every service the vendor offers today","C) The report's period ended before the assistant existed, so it does not cover the assistant. Every report has a boundary, set by its scope statement and its period. A report whose period ended before the vendor launched the assistant does not cover the assistant, however sound the platform controls were. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz" "A scheduling agent got caught in a loop overnight. It sent 400 meeting invitations and made 9,000 model calls before anyone noticed. Which control would have bounded this incident? | A) A system prompt telling the agent to stop and ask for help after a reasonable number of attempts. / B) A monitoring dashboard that shows the agent's activity, reviewed by the operations team each morning. / C) Caps on the agent's steps, spend and rate of action, enforced in code outside the model. / D) A newer and more capable model, chosen for how it reasons when a tool returns an unexpected error.","C) Caps on the agent's steps, spend and rate of action, enforced in code outside the model.. Hard limits in code hold whatever the model decides. Cap the number of steps an agent may take, the money it may spend and the rate at which it may act, and give it a kill switch. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz" "An architect proposes letting the model decide whether a user may open a record, because the user's role is written in the prompt. Under a zero trust design, which component should decide and enforce that access? | A) Identity and policy services decide, and the gateway, orchestration code and downstream systems enforce. / B) The model decides, because it reads the role and the request together and can weigh them in context. / C) The model decides, and the orchestration code enforces whatever answer the model returns. / D) The vector store decides, using the role that the model passes to it along with each query.","A) Identity and policy services decide, and the gateway, orchestration code and downstream systems enforce.. Zero trust separates the component that decides whether a request is allowed from the component that enforces the decision. In an AI system the model is neither: its context mixes your instructions with content strangers wrote, so it cannot be trusted to decide or to enforce. Your identity and policy services decide, and the gateway, the orchestration code and the downstream systems enforce. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz" "A policy assistant answers staff questions from retrieved passages. When the passages do not cover a question, it still gives a fluent, confident answer that is wrong. Which design change BEST reduces this? | A) Check that each answer is supported by the retrieved passages, and have the assistant say it does not know when support falls short. / B) Add a disclaimer under every answer saying that the assistant can make mistakes and should be checked. / C) Switch to a larger model, which is less likely to give a wrong answer when passages are missing. / D) Add an output filter that blocks any answer containing a number or a date.","A) Check that each answer is supported by the retrieved passages, and have the assistant say it does not know when support falls short.. Grounding constrains an answer to verifiable source content and checks that the content supports it. Abstention lets the system decline, or say it does not know, when support falls below a threshold. Together they act on the cause, which is an answer with nothing behind it. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz" "The alert count from your AI gateway falls by 80% in one week. The gateway team has reported no change. What should you do FIRST? | A) Report to the committee that the controls are working, since attacks have dropped. / B) Check whether the gateway is still seeing traffic, and whether its configuration has changed. / C) Tighten the gateway's rules, so that the alert count returns to its earlier level. / D) Switch the gateway off and review it at the next quarterly meeting, since it shows no value.","B) Check whether the gateway is still seeing traffic, and whether its configuration has changed.. A sudden drop in alerts can mean the attacks stopped. It can also mean the gateway stopped seeing traffic, or that someone changed or switched off its rules. Monitor the monitor: check the event stream for gaps, look for configuration changes and confirm that the tool is still receiving what it should. Source: Credential Press, https://credentialpress.com/guides/aaism-quiz"
Where next?
Also free: the AAISM interview questions and the AAISM mind map, plus every other credential on our study tools page. For the full syllabus, the AAISM Exam Guide and the AAISM Practice Questions go domain by domain.
Frequently asked questions
Is this AAISM quiz like the real exam?
No. It is a 10-question recall check written by Credential Press, not exam questions. Each answer cites its source.
Can I study the AAISM quiz offline?
Yes. Download the free flashcard CSV and import it into Anki or any flashcard app that reads CSV.
Over 6 months, your company has approved five AI pilots. Each time, business units argued about who could approve a new use case, who could accept its risk and who could stop a deployment. Which action would BEST stop the same argument from returning with the next model?
Agree an AI governance charter that settles approval, risk acceptance and stop authority once. The AI governance charter is the approved document that sets up AI oversight. It is where you settle, once, who approves new AI use cases, who accepts AI risk and who can stop a deployment. Without it, each question is argued again for every new model.
Which books go deeper on AAISM?

A study companion for the ISACA Advanced in AI Security Management (AAISM) examination. 13 chapters, 371 pages.

Practice questions with full rationales, sized to the published domain weights. 213 pages.
How to credit this page
Using a figure or a passage from this page in an article, a course, a forum answer or AI-generated text? Please credit Credential Press and link to this page, so your readers can check the source and the date it was reviewed. Copy a ready-made citation:
Plain text
Markdown
HTML
Sources
- ISACA, AAISM Exam Candidate Guide, version 1.26, with the AAISM Examination Content Outline (June 2025), checked against ISACA's published copy on 14 November 2026
- ISACA, AAISM Exam Content Outline, read 14 November 2026
- ISACA, AAISM certification page, read 16 November 2026
- ISACA, Get AAISM Certified, read 16 November 2026
Credential Press is independent of ISACA.