Skip to main content
CertRampLearning

ISACA · Free practice questions

AAISM Practice Questions

15 original ISACA AAISM practice questions across the three domains of the AAISM exam content outline, weighted roughly like the real exam. Each one comes with the answer, a short explanation and the reason every other option is wrong.

Written for the current AAISM exam · Updated October 27, 2026

These are original practice questions written by CertRamp to the official ISACA exam outline. They are not real exam questions and do not come from any exam. Try each question first, then open the answer to see the explanation for every option.

Domain 1

AI Governance and Program Management

AI asset and data life cycle managementFoundationDomain: AI Governance and Program Management

Question 1. A newly appointed AI security manager learns that several business units use AI tools, but nobody can say which ones or what data they process. What should the manager do FIRST?

  1. Option A: Block all AI services at the web proxy until a policy is approved
  2. Option B: Establish an inventory of AI systems and use cases, including their owners and the data they use
  3. Option C: Commission a penetration test of the most popular AI tool
  4. Option D: Buy an AI security monitoring product
Show answer and explanation

Correct answer: B. Establish an inventory of AI systems and use cases, including their owners and the data they use

Visibility comes first. An AI inventory that records each system, its business owner, purpose and data allows the programme to prioritise risk assessments and controls instead of reacting tool by tool.

  • A (incorrect)A blanket block disrupts the business without knowing what is in use or how critical it is, and often drives usage further into the shadows.
  • B (correct)Correct. You cannot govern, risk-assess or protect AI systems you do not know about. An inventory with owners and data flows is the basis for every later governance and risk decision.
  • C (incorrect)Testing one tool is premature when the overall landscape and its risks are unknown.
  • D (incorrect)Selecting technology before understanding the scope and requirements is a classic management error.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

AI-related strategies, policies, and proceduresIntermediateDomain: AI Governance and Program Management

Question 2. Staff are pasting confidential contract text into public generative AI chat services. Which action BEST addresses the root cause from a governance perspective?

  1. Option A: Publish an acceptable use policy for AI that defines approved tools and permitted data, supported by awareness training
  2. Option B: Discipline the employees involved
  3. Option C: Ask the public AI provider to delete the submitted data
  4. Option D: Encrypt all contracts at rest
Show answer and explanation

Correct answer: A. Publish an acceptable use policy for AI that defines approved tools and permitted data, supported by awareness training

Governance questions usually reward the answer that sets direction: policy, roles and awareness. Technical controls such as data loss prevention then enforce the policy rather than replace it.

  • A (correct)Correct. The root cause is the absence of clear rules and awareness. A policy that names approved tools and data classes, backed by training, sets the expectation that technical controls can then enforce.
  • B (incorrect)Sanctions without a published rule are hard to justify and do not prevent recurrence across the organisation.
  • C (incorrect)This may limit damage from past incidents, but it does nothing to prevent the behaviour.
  • D (incorrect)Encryption at rest does not stop authorised users from copying content into an external service.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

AI security program development and managementAdvancedDomain: AI Governance and Program Management

Question 3. The board asks how effective the AI security programme is. Which metric gives the MOST meaningful view for this audience?

  1. Option A: The number of prompts blocked by the AI gateway each month
  2. Option B: The number of AI security tools deployed
  3. Option C: The percentage of high-risk AI use cases with a completed risk assessment and approved controls
  4. Option D: The average response time of the AI models in production
Show answer and explanation

Correct answer: C. The percentage of high-risk AI use cases with a completed risk assessment and approved controls

Metrics for senior management should show whether the programme reduces risk to acceptable levels. Coverage of high-risk use cases by assessments and controls is far more meaningful to a board than raw operational counts.

  • A (incorrect)An operational activity count that says little about whether business risk is under control.
  • B (incorrect)Tool counts measure spending, not effectiveness.
  • C (correct)Correct. This metric links the programme to risk coverage of the most critical AI use cases, which is what the board needs to judge.
  • D (incorrect)A performance measure, not a security programme measure.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

Domain 2

AI Risk Management

AI risk assessment, thresholds, and treatmentIntermediateDomain: AI Risk Management

Question 4. Marketing wants to deploy a generative AI assistant that drafts customer emails. Who should be accountable for accepting the residual risk of this use case?

  1. Option A: The AI security manager
  2. Option B: The AI model provider
  3. Option C: The data scientist who configures the assistant
  4. Option D: The business owner of the marketing use case
Show answer and explanation

Correct answer: D. The business owner of the marketing use case

Good AI governance assigns clear ownership. Security identifies and assesses risks and proposes controls, but the accountable business owner decides whether residual risk is acceptable within the organisation's risk appetite.

  • A (incorrect)Security advises on risk and controls but does not own business risk decisions for a use case it does not run.
  • B (incorrect)The provider is responsible for its service, but cannot accept risk on behalf of the organisation.
  • C (incorrect)Technical staff implement the solution; they lack the authority to accept business risk.
  • D (correct)Correct. Risk acceptance sits with the business owner who benefits from the use case and has the authority to decide, informed by security's assessment.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

AI vendor and supply chain managementIntermediateDomain: AI Risk Management

Question 5. Procurement is about to sign a contract with a vendor whose product embeds a third-party large language model. Which contract provision is MOST important from an AI security management perspective?

  1. Option A: A discount for multi-year commitment
  2. Option B: Terms that restrict the use of the organisation's data for model training and require notification of security incidents
  3. Option C: A clause naming the vendor's preferred cloud region for marketing material
  4. Option D: A requirement that the vendor uses the newest model version at all times
Show answer and explanation

Correct answer: B. Terms that restrict the use of the organisation's data for model training and require notification of security incidents

Third-party AI introduces supply chain and data use risks. Contracts should cover how customer data may be used, including for training, plus incident notification, audit rights and change notification for models.

  • A (incorrect)A commercial term with no security value.
  • B (correct)Correct. These terms address the specific AI risks of the arrangement: the organisation's data being reused to train models and timely knowledge of incidents affecting it.
  • C (incorrect)Irrelevant to the security of the organisation's data.
  • D (incorrect)Automatic model changes can introduce new behaviour and risk; this is not a security safeguard.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

AI risk assessment, thresholds, and treatmentFoundationDomain: AI Risk Management

Question 6. When is the BEST time to perform the initial risk assessment for a new AI system?

  1. Option A: After the first security incident involving the system
  2. Option B: Once the system has been in production for a year
  3. Option C: During design, before data is collected and the model is built
  4. Option D: Only when an external auditor requests it
Show answer and explanation

Correct answer: C. During design, before data is collected and the model is built

Risk should be assessed at the start of the AI life cycle and revisited at major changes and periodically in operation. Early assessment is cheaper and gives decision makers real options.

  • A (incorrect)Reactive assessment leaves the organisation exposed during the period it matters most.
  • B (incorrect)By then design decisions and data choices are hard and expensive to change.
  • C (correct)Correct. Assessing risk early lets security and privacy requirements shape data selection, architecture and controls, and allows a no-go decision before significant investment.
  • D (incorrect)Risk management is an ongoing management responsibility, not an audit-driven exercise.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

AI risk assessment, thresholds, and treatmentIntermediateDomain: AI Risk Management

Question 7. A fraud detection model approved last year now misses a growing share of fraudulent transactions because customer behaviour has changed. How should this risk be handled?

  1. Option A: Treat it as model drift: monitor performance against agreed thresholds and trigger reassessment and retraining when they are breached
  2. Option B: Accept it, because the model was approved last year
  3. Option C: Report it as a data poisoning attack
  4. Option D: Replace the model with manual review of all transactions immediately
Show answer and explanation

Correct answer: A. Treat it as model drift: monitor performance against agreed thresholds and trigger reassessment and retraining when they are breached

AI risk changes after deployment. Continuous monitoring of model performance, with thresholds owned by the business, turns drift from a silent failure into a managed risk.

  • A (correct)Correct. Performance degradation as real-world data changes is drift. Ongoing monitoring with defined thresholds and a reassessment process is the appropriate risk response.
  • B (incorrect)Approval does not make a model's risk static; conditions have changed and the original assessment no longer holds.
  • C (incorrect)Poisoning is deliberate manipulation of training data. Nothing indicates an attacker here; the environment simply changed.
  • D (incorrect)A disproportionate response that may be operationally impossible; the risk can be managed through monitoring and retraining.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

AI risk assessment, thresholds, and treatmentAdvancedDomain: AI Risk Management

Question 8. An AI system that recommends loan decisions has a residual risk of unfair outcomes that exceeds the organisation's risk appetite, even after bias testing and retraining. The business still wants to use it. What is the MOST appropriate risk response?

  1. Option A: Accept the risk, because the business wants the system
  2. Option B: Transfer the risk by buying cyber insurance
  3. Option C: Ignore the issue until a regulator raises it
  4. Option D: Mitigate further, for example by requiring human review of the recommendations, and escalate the decision if risk still exceeds appetite
Show answer and explanation

Correct answer: D. Mitigate further, for example by requiring human review of the recommendations, and escalate the decision if risk still exceeds appetite

Risk within appetite can be accepted by the owner; risk above appetite needs further treatment or escalation. Human oversight is a common mitigating control for high-impact automated decisions.

  • A (incorrect)Risk above appetite cannot simply be accepted at the business's wish; that would override the organisation's own risk criteria.
  • B (incorrect)Insurance does not address unfair outcomes for customers or the regulatory and reputational consequences.
  • C (incorrect)Not a risk response at all.
  • D (correct)Correct. Additional controls such as human oversight reduce the risk; if it remains above appetite, the decision must be escalated to someone with the authority to decide.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

AI threat and vulnerability managementIntermediateDomain: AI Risk Management

Question 9. A team is threat modelling a customer support chatbot that can look up order details through an internal API. Which threat deserves the HIGHEST priority?

  1. Option A: The chatbot occasionally gives an unhelpful answer about store opening hours
  2. Option B: A user manipulates the chatbot through crafted prompts into retrieving other customers' order details
  3. Option C: The chatbot's response time increases slightly at peak hours
  4. Option D: Users dislike the chatbot's tone
Show answer and explanation

Correct answer: B. A user manipulates the chatbot through crafted prompts into retrieving other customers' order details

Prioritise threats by likelihood and impact. When a language model can call tools or APIs, prompt injection becomes a path to real data access, so it ranks far above quality or performance issues.

  • A (incorrect)A quality issue with low impact.
  • B (correct)Correct. Prompt injection combined with access to an internal API can lead to unauthorised disclosure of personal data: high likelihood for an internet-facing system and high impact.
  • C (incorrect)A performance concern, not a priority security threat.
  • D (incorrect)A user experience issue with no security impact.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

Domain 3

AI Technologies and Controls

AI life cycle (e.g., model selection, training, and validation)FoundationDomain: AI Technologies and Controls

Question 10. An attacker inserts carefully crafted records into a dataset that will later be used to train a spam filter, so that messages containing a certain phrase are classified as legitimate. What is this attack called?

  1. Option A: Data poisoning
  2. Option B: Model extraction
  3. Option C: Membership inference
  4. Option D: Denial of service
Show answer and explanation

Correct answer: A. Data poisoning

Poisoning attacks target the training stage. Controls include provenance checks on data sources, integrity protection of datasets, data validation and testing the trained model for unexpected behaviour.

  • A (correct)Correct. Poisoning corrupts training data to change how the model behaves, here by creating a hidden backdoor triggered by a phrase.
  • B (incorrect)Model extraction copies a model's behaviour by querying it; it does not change the training data.
  • C (incorrect)Membership inference tries to learn whether a specific record was in the training data.
  • D (incorrect)Denial of service targets availability, not the model's learned behaviour.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

AI security architecture and designAdvancedDomain: AI Technologies and Controls

Question 11. An AI agent reads incoming supplier emails and can create payments in the finance system. Which control MOST effectively limits the impact of a malicious instruction hidden in an email?

  1. Option A: Add a line to the system prompt telling the agent to ignore instructions found in emails
  2. Option B: Use a larger, more capable language model
  3. Option C: Restrict the agent to least-privilege permissions and require human approval before any payment is executed
  4. Option D: Log the agent's actions for monthly review
Show answer and explanation

Correct answer: C. Restrict the agent to least-privilege permissions and require human approval before any payment is executed

Indirect prompt injection cannot be ruled out entirely, so controls must limit impact. Least privilege for agent tools and human approval for irreversible or high-value actions are the most dependable safeguards.

  • A (incorrect)Prompt-level instructions can be bypassed and are not a reliable security control on their own.
  • B (incorrect)A more capable model is not inherently resistant to injected instructions.
  • C (correct)Correct. Limiting what the agent can do and adding a human approval step for high-impact actions contains the damage even if the injection succeeds.
  • D (incorrect)Logging is useful for detection, but a monthly review comes far too late to stop a fraudulent payment.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

AI life cycle (e.g., model selection, training, and validation)IntermediateDomain: AI Technologies and Controls

Question 12. Developers want to download pre-trained models from a public model repository for use in production. Which control BEST reduces the supply chain risk?

  1. Option A: Allow any model with a high number of downloads
  2. Option B: Let each developer decide which models are trustworthy
  3. Option C: Disable antivirus on build servers to avoid false positives on model files
  4. Option D: Use an approved internal model registry where models are verified for source and integrity and scanned before release
Show answer and explanation

Correct answer: D. Use an approved internal model registry where models are verified for source and integrity and scanned before release

Pre-trained models are third-party components and can carry backdoors or malicious code. Treat them like software dependencies: approved sources, integrity checks, scanning and a managed internal registry.

  • A (incorrect)Popularity is not evidence of integrity; attackers deliberately target popular names.
  • B (incorrect)Inconsistent, undocumented decisions are not a control.
  • C (incorrect)This weakens security rather than strengthening it.
  • D (correct)Correct. A controlled registry with provenance, integrity verification and scanning (for example for unsafe serialisation formats) brings external models under change and asset management.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

Security controls and monitoringIntermediateDomain: AI Technologies and Controls

Question 13. A competitor is suspected of sending large numbers of systematic queries to a public pricing-prediction API in order to recreate the model. Which control is MOST appropriate?

  1. Option A: Retrain the model on a larger dataset
  2. Option B: Rate limiting and per-client quotas with monitoring for unusual query patterns
  3. Option C: Encrypt the training data at rest
  4. Option D: Publish the model's source code to remove the incentive
Show answer and explanation

Correct answer: B. Rate limiting and per-client quotas with monitoring for unusual query patterns

Model extraction (model stealing) works through the inference interface. Access controls, rate limiting, returning only necessary output detail and monitoring of query patterns are the main defences.

  • A (incorrect)Retraining does not stop someone copying the model through its interface.
  • B (correct)Correct. Model extraction depends on high query volumes; rate limits, quotas, authentication and anomaly detection make it slower, more costly and easier to detect.
  • C (incorrect)Encryption at rest protects stored data, not a model exposed through an API.
  • D (incorrect)This would give away the asset the organisation is trying to protect.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

Data management controlsAdvancedDomain: AI Technologies and Controls

Question 14. A model will be trained on customer records that contain personal data, but the use case only needs aggregated behaviour patterns. Which control should be applied FIRST?

  1. Option A: Data minimisation: remove or de-identify personal data fields that the use case does not need before training
  2. Option B: Train the model first and filter personal data from its outputs later
  3. Option C: Make the model available only during business hours
  4. Option D: Increase the number of training epochs
Show answer and explanation

Correct answer: A. Data minimisation: remove or de-identify personal data fields that the use case does not need before training

Privacy controls are most effective early in the AI data life cycle. Minimisation and de-identification before training reduce the risk of membership inference and data leakage through model outputs.

  • A (correct)Correct. Removing unnecessary personal data at the source reduces privacy risk, including the risk of the model memorising and later revealing it, and is cheaper than compensating controls downstream.
  • B (incorrect)Output filtering is a useful additional layer, but it leaves personal data inside the model and is easier to bypass.
  • C (incorrect)Time-based restrictions do not address the presence of personal data in training.
  • D (incorrect)More training can increase memorisation; it is not a privacy control.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

Security controls and monitoringIntermediateDomain: AI Technologies and Controls

Question 15. An internal web application displays text generated by a language model directly in users' browsers. A tester shows that the model can be induced to produce script code that runs in the browser. What is the BEST fix?

  1. Option A: Ask users not to enter unusual prompts
  2. Option B: Switch to a different language model provider
  3. Option C: Treat model output as untrusted input: validate and encode it before rendering it in the browser
  4. Option D: Increase the model's temperature setting
Show answer and explanation

Correct answer: C. Treat model output as untrusted input: validate and encode it before rendering it in the browser

Improper output handling (OWASP LLM05:2025) turns model responses into classic injection flaws such as cross-site scripting. Applications should never trust model output and must apply the same validation and encoding as for user input.

  • A (incorrect)User guidance is not a technical control and does not stop a malicious user.
  • B (incorrect)Any model can be induced to produce such output; the flaw is in the application.
  • C (correct)Correct. Model output must be handled like any other untrusted data. Output encoding and validation prevent injected script from executing.
  • D (incorrect)Temperature changes randomness, not security.

Reference: ISACA: AAISM exam content outline (opens in a new tab)

Want more practice?

15 questions are a sample. The exam is 90 multiple-choice questions items.

The full CertRamp AAISM practice exams have 540 questions in 6 tests that grow with you, from a diagnostic start to an exam-level Test 4 and a challenge set beyond the exam, with an explanation for every option.

Who writes the exams

Written by a practitioner, not a question mill

Joshua Ravnjak, founder of CertRamp Learning

Joshua Ravnjak

Senior Consultant, digital transformation and AI

Founder, CertRamp Learning

921 learners and 61 reviews on Udemy (as of November 3, 2026)

I have more than 12 years of professional experience and work as a senior consultant in an IT management consultancy, focusing on digital transformation, AI and requirements engineering in the insurance industry. I have earned many certifications myself, so I know what exam preparation feels like. That is why every CertRamp practice exam follows the same progression: an easy start to see where you stand, then real exam level, then beyond it. You walk into the exam with a buffer and a good feeling.

Selected certifications

  • IREB® CPRE Foundation Level
  • ITIL® 4 Foundation
  • Professional Scrum Master I (PSM I)
  • Professional Scrum Product Owner I (PSPO I)
  • Certified SAFe® 6 Scrum Master
  • Certified AI-Empowered SAFe® 6 Product Owner/Product Manager

Plus further certifications from Microsoft and Anthropic (Claude).

ITIL® is a registered trademark of PeopleCert. IREB® is a registered trademark of IREB e.V. SAFe® is a registered trademark of Scaled Agile, Inc. PSM and PSPO are trademarks of Scrum.org.

Also in print

Cover: CPRE Foundation Level Prüfungsvorbereitung 2026 by Joshua Ravnjak

CPRE Foundation Level Prüfungsvorbereitung 2026

6 Übungsprüfungen mit 270 Fragen und ausführlichen Erklärungen

Paperback, German, October 2026

My CPRE Foundation Level exam-prep book (in German) follows the same progressive approach.

View on Amazon (opens in a new tab)