AAISM

Free AAISM practice questions

10 questions from our AAISM practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.

Back to the AAISM practice exams

Question 1 of 10, AI Governance and Program Management

Under the EU AI Act, a company plans to deploy an AI system that evaluates the creditworthiness of individual consumers. How should the governance team expect this system to be classified?

  1. AAs a limited-risk system subject only to transparency notices to the consumers
  2. BAs a high-risk system with risk management, documentation and oversight duties Correct answer
  3. CAs a minimal-risk system that is covered only by voluntary codes of conduct for providers
  4. DAs a prohibited practice that cannot be placed on the EU market at all

Answer and explanation: Creditworthiness assessment of natural persons is listed among the high-risk use cases in Annex III of the EU AI Act, triggering requirements such as a risk management system, data governance, technical documentation and human oversight. It is not one of the prohibited practices such as social scoring by public authorities. Limited-risk and minimal-risk categories apply to uses with lower potential impact on fundamental rights.

Question 2 of 10, AI Governance and Program Management

An insurer's AI security policy states that all models must be risk assessed before production use, but data science teams regularly deploy models without assessments. What should the information security manager do FIRST?

  1. AEscalate the noncompliance to the board audit committee as a material control failure
  2. BRevoke production deployment permissions for all data science teams until they demonstrate compliance
  3. CRewrite the policy to make the risk assessment requirement optional for low-value models
  4. DFind out why the requirement is bypassed, such as an unclear process or delivery pressure Correct answer

Answer and explanation: Before choosing an enforcement or policy change, the manager needs to understand the root cause of the noncompliance, which may be an unworkable process, lack of awareness or schedule pressure. Revoking permissions and escalating to the board are disproportionate first steps without facts. Weakening the policy may be appropriate later, but only after analysis shows the requirement is not risk-justified.

Question 3 of 10, AI Governance and Program Management

A hospital deploys a vendor model that prioritizes emergency department triage. The security manager is updating the business continuity plan for this service. Which consideration is MOST important to include?

  1. AA manual triage procedure staff can switch to when the model is unavailable or untrustworthy Correct answer
  2. BContract terms that require the vendor to refund fees for every hour the model is unavailable
  3. CReplicating the model to a second cloud region so that inference can fail over automatically if the primary region goes down
  4. DA target for restoring the model from backup within the recovery point objective of the hospital

Answer and explanation: For a life-safety process, continuity depends on being able to keep triaging patients when the AI output cannot be relied on, which includes cases where the model is up but behaving incorrectly. Regional failover and restore targets only address availability, not integrity failures such as poisoning or drift. Refund clauses compensate financially but do not keep the clinical process running.

Question 4 of 10, AI Risk Management

A fraud detection model's false negative rate has risen steadily over three months, although the model code and infrastructure are unchanged. The risk register rates the model's risk as low. What should the security manager do FIRST?

  1. AReplace the model with the previous version, which performed better at deployment
  2. BRetrain the model immediately on the most recent three months of transaction data to restore accuracy
  3. CIncrease the fraud alert threshold so that analysts receive fewer false alarms
  4. DTrigger a reassessment of the model's risk, as degradation is a reassessment trigger Correct answer

Answer and explanation: Degrading performance, likely from data or concept drift or possibly deliberate evasion, is a change in the risk profile, so the existing low rating is no longer reliable and must be reassessed before choosing treatment. Retraining or rolling back are possible treatments, but choosing one without analysis may embed poisoned data or ignore the cause. Raising the alert threshold would increase missed fraud.

Question 5 of 10, AI Risk Management

In the OWASP Top 10 for LLM Applications, which risk describes an attacker embedding hidden instructions in a web page that an LLM-based assistant later retrieves and follows?

  1. AExcessive agency
  2. BSystem prompt leakage
  3. CIndirect prompt injection Correct answer
  4. DData and model poisoning

Answer and explanation: Indirect prompt injection occurs when instructions arrive through external content the model processes, such as web pages, emails or documents, rather than from the user directly. Data poisoning corrupts training or fine-tuning data rather than runtime input. System prompt leakage concerns exposure of the hidden instructions, and unbounded consumption concerns resource exhaustion and cost abuse.

Question 6 of 10, AI Risk Management

A risk workshop is developing AI risk scenarios for a customer onboarding assistant. Which scenario description is MOST useful for assessment?

  1. APrompt injection is listed in the OWASP Top 10 for LLM Applications, so the onboarding assistant is generally exposed to that category of threat
  2. BThere is a risk that the assistant may not meet customer expectations
  3. CThe assistant could be attacked using a variety of AI-specific techniques, which may lead to a range of negative consequences for the organization
  4. DAn attacker hides instructions in an uploaded document so the assistant reveals other customers' data, causing regulatory penalties Correct answer

Answer and explanation: A useful risk scenario names the threat actor, the method, the asset affected and the business impact, so likelihood and impact can be estimated and controls mapped. Citing a threat category from OWASP, saying the assistant could be attacked in various ways, or worrying about unmet expectations are too generic to assess or treat because they omit the actor, asset or consequence.

Question 7 of 10, AI Technologies and Controls

A data science team wants to train a model on customer transaction records and publish aggregate insights, while limiting the risk that any individual's participation can be inferred from the model. Which technique BEST meets this requirement?

  1. AEncrypting the training dataset with AES-256 while it is stored in the data lake
  2. BTraining the model in an isolated virtual network with no internet egress and tightly restricted analyst access
  3. CRemoving names and account numbers from the records before training begins
  4. DTraining with differential privacy so each record's influence on the model is bounded Correct answer

Answer and explanation: Differential privacy adds calibrated noise during training so that the presence or absence of any single record has a mathematically bounded effect, directly reducing membership inference risk. Removing direct identifiers still allows inference from the remaining attributes. Encryption at rest and network isolation protect the data during storage and training, not against inference from the trained model.

Question 8 of 10, AI Technologies and Controls

A company is building a retrieval-augmented generation (RAG) assistant over its document repositories, which have different access permissions. Which control BEST prevents the assistant from revealing documents a user is not entitled to see?

  1. AFilter retrieved chunks against the requesting user's permissions before generation Correct answer
  2. BInstruct the model in its system prompt never to disclose confidential documents
  3. CEncrypt the vector database at rest with customer-managed keys held in a hardware security module
  4. DRestrict the assistant to employees who have completed data handling training

Answer and explanation: Permission-aware retrieval enforces existing access rights at query time so that the model only receives content the user is already authorized to read. System prompt instructions can be bypassed through prompt injection and are not an access control. Encryption at rest and training do not stop an authorized user of the assistant from receiving another user's documents.

Question 9 of 10, AI Technologies and Controls

An AI agent can read emails, search the web and send payments through an internal API to automate supplier invoicing. Which control BEST addresses the risk of excessive agency?

  1. ARunning the agent only during business hours when finance staff can watch its activity logs and stop it if needed
  2. BUsing the most capable available foundation model so that the agent makes fewer reasoning mistakes
  3. CAdding a system prompt instruction that tells the agent to ignore instructions contained in emails
  4. DRequiring human approval for payments and scoping the agent's API credentials to minimum actions Correct answer

Answer and explanation: Excessive agency is mitigated by limiting the agent's permissions and functions to what is necessary and requiring human confirmation for high-impact actions such as payments. A more capable model can still be manipulated by injected instructions. Prompt instructions are not a reliable boundary, and time-of-day restrictions do not prevent an unauthorized payment.

Question 10 of 10, AI Technologies and Controls

Labels for a content moderation dataset show low agreement between annotators on what counts as harassment. What is the MOST appropriate data management response?

  1. AClarify guidelines, retrain annotators and measure agreement before use Correct answer
  2. BUse the labels as they are, because the model will learn to resolve disagreements
  3. CDiscard the harassment category entirely from the moderation model
  4. DKeep only labels from the fastest annotator, since speed reflects experience

Answer and explanation: Low inter-annotator agreement signals ambiguous guidelines or inconsistent training, which produces unreliable and potentially biased labels; improving guidelines and measuring agreement fixes the data at the source. The model cannot resolve inconsistent ground truth, speed does not indicate accuracy, and dropping the category removes needed protection.