IAPP AIGP Practice Test 83
Question 820
An organization exposes a model endpoint and retrieval store without considering training data poisoning or model extraction. Model assurance reviewers need a documented control before approving continued experimentation. What is the most relevant fallback when the promised safeguard cannot be demonstrated?
- Restore the last valid feature pipeline or model while investigating the failure.
- Rate-limit, revoke and isolate compromised components while preserving forensic evidence.
- Suspend onboarding where missing disclosures would make safe deployment impossible.
- Roll back to a known-good model and isolate faulty changes after adverse signals.
Explanation
Correct Answer
B:Rate-limit, revoke and isolate compromised components while preserving forensic evidence. Correct. It limits further reliance or exposure while the risk remains unresolved. AI security includes attacks against data, models, prompts and inference services.
Incorrect Answers
A:Restore the last valid feature pipeline or model while investigating the failure. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Root-cause analysis of model failures: AI root-cause analysis must distinguish data drift, code defects, feedback loops and model faults. The present case instead concerns AI security threat modeling.
C:Suspend onboarding where missing disclosures would make safe deployment impossible. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Public and downstream disclosure: Developer-to-deployer documentation transfers essential risk information across the AI supply chain. The present case instead concerns AI security threat modeling.
D:Roll back to a known-good model and isolate faulty changes after adverse signals. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Version deprecation and rollback: Model lifecycle governance needs controlled retirement and rollback, not irreversible swaps. The present case instead concerns AI security threat modeling.
Question 821
An organization exposes a model endpoint and retrieval store without considering training data poisoning or model extraction. A material training or architecture change makes the previous technical approval insufficient. What review step identifies the specific control gap rather than an unrelated AI risk?
- Inspect retraining triggers, dataset changes, performance evidence and promotion permissions.
- Map threat actors, exposed interfaces, model assets, data and attack paths.
- Identify missing model facts including training scope, intended use, evaluation, tradeoffs and limitations.
- Check audit independence, scope, criteria, sampling and access to underlying evidence.
Explanation
Correct Answer
B:Map threat actors, exposed interfaces, model assets, data and attack paths. Correct. It establishes the relevant factual risk boundary before implementation. AI security includes attacks against data, models, prompts and inference services.
Incorrect Answers
A:Inspect retraining triggers, dataset changes, performance evidence and promotion permissions. Not selected. That check primarily diagnoses a different governance issue. It belongs with Scheduled retraining governance: Retraining is a new model change that requires testing, traceability and accountable release. The present case instead concerns AI security threat modeling.
C:Identify missing model facts including training scope, intended use, evaluation, tradeoffs and limitations. Not selected. That check primarily diagnoses a different governance issue. It belongs with Model cards and limitations: Model cards communicate evidence, limits and proper interpretation to deployment teams. The present case instead concerns AI security threat modeling.
D:Check audit independence, scope, criteria, sampling and access to underlying evidence. Not selected. That check primarily diagnoses a different governance issue. It belongs with Periodic independent audits: Independent audit contributes objective scrutiny that ordinary self-monitoring cannot replace. The present case instead concerns AI security threat modeling.
Question 822
An organization exposes a model endpoint and retrieval store without considering training data poisoning or model extraction. A material training or architecture change makes the previous technical approval insufficient. Which safeguard most directly treats the described failure mode?
- Apply security controls for model APIs, training pipelines, retrieval and privileged operations.
- Supply accurate instructions, risk information, performance limits and post-market monitoring contacts.
- Require controlled retraining with versioned data, independent evaluation and approval before deployment.
- Run structured cross-functional causal analysis before changing thresholds or retraining blindly.
Explanation
Correct Answer
A:Apply security controls for model APIs, training pipelines, retrieval and privileged operations. Correct. It directly reduces or governs the specific documented risk. AI security includes attacks against data, models, prompts and inference services.
Incorrect Answers
B:Supply accurate instructions, risk information, performance limits and post-market monitoring contacts. Not selected. That action treats another control objective, not the specific gap described. It belongs with Public and downstream disclosure: Developer-to-deployer documentation transfers essential risk information across the AI supply chain. The present case instead concerns AI security threat modeling.
C:Require controlled retraining with versioned data, independent evaluation and approval before deployment. Not selected. That action treats another control objective, not the specific gap described. It belongs with Scheduled retraining governance: Retraining is a new model change that requires testing, traceability and accountable release. The present case instead concerns AI security threat modeling.
D:Run structured cross-functional causal analysis before changing thresholds or retraining blindly. Not selected. That action treats another control objective, not the specific gap described. It belongs with Root-cause analysis of model failures: AI root-cause analysis must distinguish data drift, code defects, feedback loops and model faults. The present case instead concerns AI security threat modeling.
Question 823
An organization exposes a model endpoint and retrieval store without considering training data poisoning or model extraction. A material training or architecture change makes the previous technical approval insufficient. Which verification task should support the next approval decision?
- Audit the release packet and verify every required decision is complete and current.
- Test access control, suspicious query detection and poisoning resistance in controlled exercises.
- Replay representative failing cases against prior and current pipeline versions.
- Compare candidate and current models on safety and subgroup test suites before promotion.
Explanation
Correct Answer
B:Test access control, suspicious query detection and poisoning resistance in controlled exercises. Correct. It yields measurable evidence that the case-specific safeguard works. AI security includes attacks against data, models, prompts and inference services.
Incorrect Answers
A:Audit the release packet and verify every required decision is complete and current. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Release readiness checklist: A release gate is a risk acceptance decision supported by evidence, not a calendar milestone. The present case instead concerns AI security threat modeling.
C:Replay representative failing cases against prior and current pipeline versions. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Root-cause analysis of model failures: AI root-cause analysis must distinguish data drift, code defects, feedback loops and model faults. The present case instead concerns AI security threat modeling.
D:Compare candidate and current models on safety and subgroup test suites before promotion. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Scheduled retraining governance: Retraining is a new model change that requires testing, traceability and accountable release. The present case instead concerns AI security threat modeling.
Question 824
An organization exposes a model endpoint and retrieval store without considering training data poisoning or model extraction. A material training or architecture change makes the previous technical approval insufficient. Select TWO complementary actions: first assess the stated concern and then implement its targeted control.
- Apply security controls for model APIs, training pipelines, retrieval and privileged operations.
- Check audit independence, scope, criteria, sampling and access to underlying evidence.
- Map threat actors, exposed interfaces, model assets, data and attack paths.
- Implement data, concept and performance drift monitoring with ownership and trigger thresholds.
Explanation
Correct Answers
A:Apply security controls for model APIs, training pipelines, retrieval and privileged operations. Correct. As the second, implementation half of the requested pair, this directly applies the needed safeguard. AI security includes attacks against data, models, prompts and inference services.
C:Map threat actors, exposed interfaces, model assets, data and attack paths. Correct. As the first, evidence-gathering half of the requested pair, this establishes the specific risk boundary. AI security includes attacks against data, models, prompts and inference services.
Incorrect Answers
B:Check audit independence, scope, criteria, sampling and access to underlying evidence. Not selected. It addresses a different governance mechanism, not the required assessment-and-control pair. It belongs with Periodic independent audits: Independent audit contributes objective scrutiny that ordinary self-monitoring cannot replace. The present case instead concerns AI security threat modeling.
D:Implement data, concept and performance drift monitoring with ownership and trigger thresholds. Not selected. It addresses a different governance mechanism, not the required assessment-and-control pair. It belongs with Monitoring and drift: AI monitoring must measure model behavior and data context, not only infrastructure availability. The present case instead concerns AI security threat modeling.
Question 825
An assistant starts generating fabricated account balances, but the on-call engineer sees no infrastructure error. The development team must show evidence before promoting the next model version. What review step identifies the specific control gap rather than an unrelated AI risk?
- Map threat actors, exposed interfaces, model assets, data and attack paths.
- Identify realistic misuse paths, privileged tools, high-risk prompts and known failure classes.
- Assess business impact, decision reliance, likelihood of harm and the anomaly's spread.
- Review release artifacts, version pins, rollback dependencies and user communication plans.
Explanation
Correct Answer
C:Assess business impact, decision reliance, likelihood of harm and the anomaly's spread. Correct. It establishes the relevant factual risk boundary before implementation. A system can be operationally available while causing a serious output-quality incident.
Incorrect Answers
A:Map threat actors, exposed interfaces, model assets, data and attack paths. Not selected. That check primarily diagnoses a different governance issue. It belongs with AI security threat modeling: AI security includes attacks against data, models, prompts and inference services. The present case instead concerns AI incident classification.
B:Identify realistic misuse paths, privileged tools, high-risk prompts and known failure classes. Not selected. That check primarily diagnoses a different governance issue. It belongs with Red-team exercise design: Adversarial testing probes how a system can fail under purposeful misuse rather than normal usage. The present case instead concerns AI incident classification.
D:Review release artifacts, version pins, rollback dependencies and user communication plans. Not selected. That check primarily diagnoses a different governance issue. It belongs with Version deprecation and rollback: Model lifecycle governance needs controlled retirement and rollback, not irreversible swaps. The present case instead concerns AI incident classification.
Question 826
An assistant starts generating fabricated account balances, but the on-call engineer sees no infrastructure error. The development team must show evidence before promoting the next model version. Which safeguard most directly treats the described failure mode?
- Create an AI-specific incident category with escalation to business owners and risk management.
- Require controlled retraining with versioned data, independent evaluation and approval before deployment.
- Schedule risk-based audits with appropriate independent challenge and documented findings management.
- Run structured cross-functional causal analysis before changing thresholds or retraining blindly.
Explanation
Correct Answer
A:Create an AI-specific incident category with escalation to business owners and risk management. Correct. It directly reduces or governs the specific documented risk. A system can be operationally available while causing a serious output-quality incident.
Incorrect Answers
B:Require controlled retraining with versioned data, independent evaluation and approval before deployment. Not selected. That action treats another control objective, not the specific gap described. It belongs with Scheduled retraining governance: Retraining is a new model change that requires testing, traceability and accountable release. The present case instead concerns AI incident classification.
C:Schedule risk-based audits with appropriate independent challenge and documented findings management. Not selected. That action treats another control objective, not the specific gap described. It belongs with Periodic independent audits: Independent audit contributes objective scrutiny that ordinary self-monitoring cannot replace. The present case instead concerns AI incident classification.
D:Run structured cross-functional causal analysis before changing thresholds or retraining blindly. Not selected. That action treats another control objective, not the specific gap described. It belongs with Root-cause analysis of model failures: AI root-cause analysis must distinguish data drift, code defects, feedback loops and model faults. The present case instead concerns AI incident classification.
Question 827
An assistant starts generating fabricated account balances, but the on-call engineer sees no infrastructure error. The development team must show evidence before promoting the next model version. Which verification task should support the next approval decision?
- Confirm the integrator can apply the necessary restrictions and escalation procedures.
- Check downstream integrators can distinguish validated use from unsupported extensions.
- Verify identified nonconformities have owners, evidence-based closure and deadlines.
- Run a tabletop test for detecting hallucination patterns and notifying affected teams.
Explanation
Correct Answer
D:Run a tabletop test for detecting hallucination patterns and notifying affected teams. Correct. It yields measurable evidence that the case-specific safeguard works. A system can be operationally available while causing a serious output-quality incident.
Incorrect Answers
A:Confirm the integrator can apply the necessary restrictions and escalation procedures. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Public and downstream disclosure: Developer-to-deployer documentation transfers essential risk information across the AI supply chain. The present case instead concerns AI incident classification.
B:Check downstream integrators can distinguish validated use from unsupported extensions. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Model cards and limitations: Model cards communicate evidence, limits and proper interpretation to deployment teams. The present case instead concerns AI incident classification.
C:Verify identified nonconformities have owners, evidence-based closure and deadlines. Not selected. That test would not demonstrate this particular risk has been addressed. It belongs with Periodic independent audits: Independent audit contributes objective scrutiny that ordinary self-monitoring cannot replace. The present case instead concerns AI incident classification.
Question 828
An assistant starts generating fabricated account balances, but the on-call engineer sees no infrastructure error. The development team must show evidence before promoting the next model version. If the control fails, which immediate response most appropriately limits the documented risk?
- Disable inaccurate financial-answer features and provide verified manual information during remediation.
- Disable vulnerable functions and escalate material findings until fixes are validated.
- Restrict unvalidated downstream claims until model documentation is corrected.
- Escalate unresolved critical findings and restrict risky releases until assurance gaps close.
Explanation
Correct Answer
A:Disable inaccurate financial-answer features and provide verified manual information during remediation. Correct. It limits further reliance or exposure while the risk remains unresolved. A system can be operationally available while causing a serious output-quality incident.
Incorrect Answers
B:Disable vulnerable functions and escalate material findings until fixes are validated. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Red-team exercise design: Adversarial testing probes how a system can fail under purposeful misuse rather than normal usage. The present case instead concerns AI incident classification.
C:Restrict unvalidated downstream claims until model documentation is corrected. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Model cards and limitations: Model cards communicate evidence, limits and proper interpretation to deployment teams. The present case instead concerns AI incident classification.
D:Escalate unresolved critical findings and restrict risky releases until assurance gaps close. Not selected. That response contains a different failure mode and leaves the stated risk unresolved. It belongs with Periodic independent audits: Independent audit contributes objective scrutiny that ordinary self-monitoring cannot replace. The present case instead concerns AI incident classification.
Question 829
An assistant starts generating fabricated account balances, but the on-call engineer sees no infrastructure error. Model assurance reviewers need a documented control before approving continued experimentation. Which initial evidence check most directly addresses the governance concern?
- Correlate model, dataset, feature pipeline and upstream service versions with the first failure time.
- Assess business impact, decision reliance, likelihood of harm and the anomaly's spread.
- Identify missing model facts including training scope, intended use, evaluation, tradeoffs and limitations.
- Inspect retraining triggers, dataset changes, performance evidence and promotion permissions.
Explanation
Correct Answer
B:Assess business impact, decision reliance, likelihood of harm and the anomaly's spread. Correct. It establishes the relevant factual risk boundary before implementation. A system can be operationally available while causing a serious output-quality incident.
Incorrect Answers
A:Correlate model, dataset, feature pipeline and upstream service versions with the first failure time. Not selected. That check primarily diagnoses a different governance issue. It belongs with Root-cause analysis of model failures: AI root-cause analysis must distinguish data drift, code defects, feedback loops and model faults. The present case instead concerns AI incident classification.
C:Identify missing model facts including training scope, intended use, evaluation, tradeoffs and limitations. Not selected. That check primarily diagnoses a different governance issue. It belongs with Model cards and limitations: Model cards communicate evidence, limits and proper interpretation to deployment teams. The present case instead concerns AI incident classification.
D:Inspect retraining triggers, dataset changes, performance evidence and promotion permissions. Not selected. That check primarily diagnoses a different governance issue. It belongs with Scheduled retraining governance: Retraining is a new model change that requires testing, traceability and accountable release. The present case instead concerns AI incident classification.
Question 830
An assistant starts generating fabricated account balances, but the on-call engineer sees no infrastructure error. Model assurance reviewers need a documented control before approving continued experimentation. Which specific implementation best addresses the identified gap?
- Schedule risk-based audits with appropriate independent challenge and documented findings management.
- Run structured cross-functional causal analysis before changing thresholds or retraining blindly.
- Maintain staged rollout, backward-compatible versioning and an exercised emergency rollback path.
- Create an AI-specific incident category with escalation to business owners and risk management.
Explanation
Correct Answer
D:Create an AI-specific incident category with escalation to business owners and risk management. Correct. It directly reduces or governs the specific documented risk. A system can be operationally available while causing a serious output-quality incident.
Incorrect Answers
A:Schedule risk-based audits with appropriate independent challenge and documented findings management. Not selected. That action treats another control objective, not the specific gap described. It belongs with Periodic independent audits: Independent audit contributes objective scrutiny that ordinary self-monitoring cannot replace. The present case instead concerns AI incident classification.
B:Run structured cross-functional causal analysis before changing thresholds or retraining blindly. Not selected. That action treats another control objective, not the specific gap described. It belongs with Root-cause analysis of model failures: AI root-cause analysis must distinguish data drift, code defects, feedback loops and model faults. The present case instead concerns AI incident classification.
C:Maintain staged rollout, backward-compatible versioning and an exercised emergency rollback path. Not selected. That action treats another control objective, not the specific gap described. It belongs with Version deprecation and rollback: Model lifecycle governance needs controlled retirement and rollback, not irreversible swaps. The present case instead concerns AI incident classification.