ISACA Advanced in AI Audit AAIA Practice Test 35
Question 340
A recommendation model changes behavior after malicious reviews enter a feedback stream. From an AI audit perspective, which corrective action is the MOST direct response to the weakness described, rather than a useful control for some different AI risk?
- Restrict external tool destinations and inspect sensitive content before egress
- Verify provenance and execute untrusted packages in restricted build environments
- Authenticate data feeds, detect anomalous submissions and quarantine suspicious updates
- Isolate untrusted content and enforce tool-level authorization and egress checks
Explanation
Correct Answer
C:Correct. Authenticate data feeds, detect anomalous submissions and quarantine suspicious updates is aimed at the actual failure: feedback data lacks integrity and poisoning detection. The audit should subsequently obtain quarantine decisions and signed provenance for accepted feedback to test execution, not merely the existence of a policy.
Incorrect Answers
A:“Restrict external tool destinations and inspect sensitive content before egress” is a control for tool egress and data-classification checks are missing, but the audit finding here is feedback data lacks integrity and poisoning detection. The relevant measure is “Authenticate data feeds, detect anomalous submissions and quarantine suspicious updates”.
B:“Verify provenance and execute untrusted packages in restricted build environments” is a control for model supply-chain files were not scanned or sandboxed before execution, but the audit finding here is feedback data lacks integrity and poisoning detection. The relevant measure is “Authenticate data feeds, detect anomalous submissions and quarantine suspicious updates”.
D:“Isolate untrusted content and enforce tool-level authorization and egress checks” is a control for retrieved content is allowed to cross the instruction trust boundary, but the audit finding here is feedback data lacks integrity and poisoning detection. The relevant measure is “Authenticate data feeds, detect anomalous submissions and quarantine suspicious updates”.
Question 341
In an audit of Threats and Vulnerabilities Specific to AI, managers say this measure is live: Authenticate data feeds, detect anomalous submissions and quarantine suspicious updates. They claim it corrected a material audit gap. What direct operating evidence would allow the auditor to substantiate that particular claim across a defensible sample?
- Proxy block logs and tested denied tool requests containing secret strings
- Quarantine decisions and signed provenance for accepted feedback
- SBOM review, sandbox network trace and blocked installer execution
- Red-team poisoned-document results and blocked file-export calls
Explanation
Correct Answer
B:Correct. The proposed records—quarantine decisions and signed provenance for accepted feedback—are directly tied to the asserted safeguard and allow its actual operation to be corroborated. The gap being tested is feedback data lacks integrity and poisoning detection.
Incorrect Answers
A:“Proxy block logs and tested denied tool requests containing secret strings” would test tool egress and data-classification checks are missing, not the present safeguard against feedback data lacks integrity and poisoning detection. The expected audit evidence is “Quarantine decisions and signed provenance for accepted feedback”.
C:“SBOM review, sandbox network trace and blocked installer execution” would test model supply-chain files were not scanned or sandboxed before execution, not the present safeguard against feedback data lacks integrity and poisoning detection. The expected audit evidence is “Quarantine decisions and signed provenance for accepted feedback”.
D:“Red-team poisoned-document results and blocked file-export calls” would test retrieved content is allowed to cross the instruction trust boundary, not the present safeguard against feedback data lacks integrity and poisoning detection. The expected audit evidence is “Quarantine decisions and signed provenance for accepted feedback”.
Question 342
In a review of Threats and Vulnerabilities Specific to AI, investigators cannot establish the expected control proof: quarantine decisions and signed provenance for accepted feedback. The underlying event was that a recommendation model changes behavior after malicious reviews enter a feedback stream. Which control weakness is the MOST likely cause requiring focused audit work?
- Feedback data lacks integrity and poisoning detection
- Tool egress and data-classification checks are missing
- Model supply-chain files were not scanned or sandboxed before execution
- Retrieved content is allowed to cross the instruction trust boundary
Explanation
Correct Answer
A:Correct. The observed condition specifically indicates feedback data lacks integrity and poisoning detection. A meaningful audit response is to verify this diagnosis and then recommend authenticate data feeds, detect anomalous submissions and quarantine suspicious updates.
Incorrect Answers
B:“Tool egress and data-classification checks are missing” concerns a separate audit weakness, not the missing assurance described by feedback data lacks integrity and poisoning detection. The correct cause is “Feedback data lacks integrity and poisoning detection”.
C:“Model supply-chain files were not scanned or sandboxed before execution” concerns a separate audit weakness, not the missing assurance described by feedback data lacks integrity and poisoning detection. The correct cause is “Feedback data lacks integrity and poisoning detection”.
D:“Retrieved content is allowed to cross the instruction trust boundary” concerns a separate audit weakness, not the missing assurance described by feedback data lacks integrity and poisoning detection. The correct cause is “Feedback data lacks integrity and poisoning detection”.
Question 343
An attacker edits model weights in object storage but release images remain correctly signed. From an AI audit perspective, which corrective action is the MOST direct response to the weakness described, rather than a useful control for some different AI risk?
- Treat OCR/image content as untrusted task data and constrain allowable actions
- Encode untrusted model output and apply strict rendering/content security controls
- Apply rate limits, anomaly detection and output disclosure controls
- Sign and verify model weight hashes at promotion and load
Explanation
Correct Answer
D:Correct. Sign and verify model weight hashes at promotion and load is aimed at the actual failure: model artifacts themselves lack integrity controls independent of container signing. The audit should subsequently obtain failed hash validation on tampered model artifact in a release test to test execution, not merely the existence of a policy.
Incorrect Answers
A:“Treat OCR/image content as untrusted task data and constrain allowable actions” is a control for visual text from documents was promoted to trusted system instruction, but the audit finding here is model artifacts themselves lack integrity controls independent of container signing. The relevant measure is “Sign and verify model weight hashes at promotion and load”.
B:“Encode untrusted model output and apply strict rendering/content security controls” is a control for model output is being trusted as executable presentation code, but the audit finding here is model artifacts themselves lack integrity controls independent of container signing. The relevant measure is “Sign and verify model weight hashes at promotion and load”.
C:“Apply rate limits, anomaly detection and output disclosure controls” is a control for query-rate and extraction defenses are inadequate for model privacy, but the audit finding here is model artifacts themselves lack integrity controls independent of container signing. The relevant measure is “Sign and verify model weight hashes at promotion and load”.
Question 344
In an audit of Threats and Vulnerabilities Specific to AI, managers say this measure is live: Sign and verify model weight hashes at promotion and load. They claim it corrected a material audit gap. What direct operating evidence would allow the auditor to substantiate that particular claim across a defensible sample?
- Adversarial invoice tests showing extraction without instruction execution
- XSS test payload blocked by renderer and browser policy
- Failed hash validation on tampered model artifact in a release test
- Extractive attack simulation, query telemetry and throttling events
Explanation
Correct Answer
C:Correct. The proposed records—failed hash validation on tampered model artifact in a release test—are directly tied to the asserted safeguard and allow its actual operation to be corroborated. The gap being tested is model artifacts themselves lack integrity controls independent of container signing.
Incorrect Answers
A:“Adversarial invoice tests showing extraction without instruction execution” would test visual text from documents was promoted to trusted system instruction, not the present safeguard against model artifacts themselves lack integrity controls independent of container signing. The expected audit evidence is “Failed hash validation on tampered model artifact in a release test”.
B:“XSS test payload blocked by renderer and browser policy” would test model output is being trusted as executable presentation code, not the present safeguard against model artifacts themselves lack integrity controls independent of container signing. The expected audit evidence is “Failed hash validation on tampered model artifact in a release test”.
D:“Extractive attack simulation, query telemetry and throttling events” would test query-rate and extraction defenses are inadequate for model privacy, not the present safeguard against model artifacts themselves lack integrity controls independent of container signing. The expected audit evidence is “Failed hash validation on tampered model artifact in a release test”.
Question 345
In a review of Threats and Vulnerabilities Specific to AI, investigators cannot establish the expected control proof: failed hash validation on tampered model artifact in a release test. The underlying event was that an attacker edits model weights in object storage but release images remain correctly signed. Which control weakness is the MOST likely cause requiring focused audit work?
- Visual text from documents was promoted to trusted system instruction
- Model artifacts themselves lack integrity controls independent of container signing
- Model output is being trusted as executable presentation code
- Query-rate and extraction defenses are inadequate for model privacy
Explanation
Correct Answer
B:Correct. The observed condition specifically indicates model artifacts themselves lack integrity controls independent of container signing. A meaningful audit response is to verify this diagnosis and then recommend sign and verify model weight hashes at promotion and load.
Incorrect Answers
A:“Visual text from documents was promoted to trusted system instruction” concerns a separate audit weakness, not the missing assurance described by model artifacts themselves lack integrity controls independent of container signing. The correct cause is “Model artifacts themselves lack integrity controls independent of container signing”.
C:“Model output is being trusted as executable presentation code” concerns a separate audit weakness, not the missing assurance described by model artifacts themselves lack integrity controls independent of container signing. The correct cause is “Model artifacts themselves lack integrity controls independent of container signing”.
D:“Query-rate and extraction defenses are inadequate for model privacy” concerns a separate audit weakness, not the missing assurance described by model artifacts themselves lack integrity controls independent of container signing. The correct cause is “Model artifacts themselves lack integrity controls independent of container signing”.
Question 346
A retrieval assistant sends confidential passages to an external search service through a hidden plugin. From an AI audit perspective, which corrective action is the MOST direct response to the weakness described, rather than a useful control for some different AI risk?
- Restrict external tool destinations and inspect sensitive content before egress
- Test membership inference and reduce memorization or public access as warranted
- Enforce server-side tenant authorization on every vector search
- Authenticate data feeds, detect anomalous submissions and quarantine suspicious updates
Explanation
Correct Answer
A:Correct. Restrict external tool destinations and inspect sensitive content before egress is aimed at the actual failure: tool egress and data-classification checks are missing. The audit should subsequently obtain proxy block logs and tested denied tool requests containing secret strings to test execution, not merely the existence of a policy.
Incorrect Answers
B:“Test membership inference and reduce memorization or public access as warranted” is a control for membership-inference exposure was not included in privacy threat assessment, but the audit finding here is tool egress and data-classification checks are missing. The relevant measure is “Restrict external tool destinations and inspect sensitive content before egress”.
C:“Enforce server-side tenant authorization on every vector search” is a control for tenant isolation is enforced only by client-provided retrieval hints, but the audit finding here is tool egress and data-classification checks are missing. The relevant measure is “Restrict external tool destinations and inspect sensitive content before egress”.
D:“Authenticate data feeds, detect anomalous submissions and quarantine suspicious updates” is a control for feedback data lacks integrity and poisoning detection, but the audit finding here is tool egress and data-classification checks are missing. The relevant measure is “Restrict external tool destinations and inspect sensitive content before egress”.
Question 347
In an audit of Threats and Vulnerabilities Specific to AI, managers say this measure is live: Restrict external tool destinations and inspect sensitive content before egress. They claim it corrected a material audit gap. What direct operating evidence would allow the auditor to substantiate that particular claim across a defensible sample?
- Privacy attack baseline, mitigation results and endpoint authorization tests
- Cross-tenant negative queries and datastore policy test results
- Quarantine decisions and signed provenance for accepted feedback
- Proxy block logs and tested denied tool requests containing secret strings
Explanation
Correct Answer
D:Correct. The proposed records—proxy block logs and tested denied tool requests containing secret strings—are directly tied to the asserted safeguard and allow its actual operation to be corroborated. The gap being tested is tool egress and data-classification checks are missing.
Incorrect Answers
A:“Privacy attack baseline, mitigation results and endpoint authorization tests” would test membership-inference exposure was not included in privacy threat assessment, not the present safeguard against tool egress and data-classification checks are missing. The expected audit evidence is “Proxy block logs and tested denied tool requests containing secret strings”.
B:“Cross-tenant negative queries and datastore policy test results” would test tenant isolation is enforced only by client-provided retrieval hints, not the present safeguard against tool egress and data-classification checks are missing. The expected audit evidence is “Proxy block logs and tested denied tool requests containing secret strings”.
C:“Quarantine decisions and signed provenance for accepted feedback” would test feedback data lacks integrity and poisoning detection, not the present safeguard against tool egress and data-classification checks are missing. The expected audit evidence is “Proxy block logs and tested denied tool requests containing secret strings”.
Question 348
In a review of Threats and Vulnerabilities Specific to AI, investigators cannot establish the expected control proof: proxy block logs and tested denied tool requests containing secret strings. The underlying event was that a retrieval assistant sends confidential passages to an external search service through a hidden plugin. Which control weakness is the MOST likely cause requiring focused audit work?
- Membership-inference exposure was not included in privacy threat assessment
- Tenant isolation is enforced only by client-provided retrieval hints
- Tool egress and data-classification checks are missing
- Feedback data lacks integrity and poisoning detection
Explanation
Correct Answer
C:Correct. The observed condition specifically indicates tool egress and data-classification checks are missing. A meaningful audit response is to verify this diagnosis and then recommend restrict external tool destinations and inspect sensitive content before egress.
Incorrect Answers
A:“Membership-inference exposure was not included in privacy threat assessment” concerns a separate audit weakness, not the missing assurance described by tool egress and data-classification checks are missing. The correct cause is “Tool egress and data-classification checks are missing”.
B:“Tenant isolation is enforced only by client-provided retrieval hints” concerns a separate audit weakness, not the missing assurance described by tool egress and data-classification checks are missing. The correct cause is “Tool egress and data-classification checks are missing”.
D:“Feedback data lacks integrity and poisoning detection” concerns a separate audit weakness, not the missing assurance described by tool egress and data-classification checks are missing. The correct cause is “Tool egress and data-classification checks are missing”.
Question 349
A multimodal assistant follows instructions embedded in a scanned invoice instead of extracting only invoice fields. From an AI audit perspective, which corrective action is the MOST direct response to the weakness described, rather than a useful control for some different AI risk?
- Verify provenance and execute untrusted packages in restricted build environments
- Treat OCR/image content as untrusted task data and constrain allowable actions
- Add adversarial interaction monitoring, guardrail hardening and escalation
- Sign and verify model weight hashes at promotion and load
Explanation
Correct Answer
B:Correct. Treat OCR/image content as untrusted task data and constrain allowable actions is aimed at the actual failure: visual text from documents was promoted to trusted system instruction. The audit should subsequently obtain adversarial invoice tests showing extraction without instruction execution to test execution, not merely the existence of a policy.
Incorrect Answers
A:“Verify provenance and execute untrusted packages in restricted build environments” is a control for model supply-chain files were not scanned or sandboxed before execution, but the audit finding here is visual text from documents was promoted to trusted system instruction. The relevant measure is “Treat OCR/image content as untrusted task data and constrain allowable actions”.
C:“Add adversarial interaction monitoring, guardrail hardening and escalation” is a control for adaptive jailbreak testing and abuse monitoring were inadequate, but the audit finding here is visual text from documents was promoted to trusted system instruction. The relevant measure is “Treat OCR/image content as untrusted task data and constrain allowable actions”.
D:“Sign and verify model weight hashes at promotion and load” is a control for model artifacts themselves lack integrity controls independent of container signing, but the audit finding here is visual text from documents was promoted to trusted system instruction. The relevant measure is “Treat OCR/image content as untrusted task data and constrain allowable actions”.
Question 350
In an audit of Threats and Vulnerabilities Specific to AI, managers say this measure is live: Treat OCR/image content as untrusted task data and constrain allowable actions. They claim it corrected a material audit gap. What direct operating evidence would allow the auditor to substantiate that particular claim across a defensible sample?
- Adversarial invoice tests showing extraction without instruction execution
- SBOM review, sandbox network trace and blocked installer execution
- Red-team adaptive prompt series and risk alert response evidence
- Failed hash validation on tampered model artifact in a release test
Explanation
Correct Answer
A:Correct. The proposed records—adversarial invoice tests showing extraction without instruction execution—are directly tied to the asserted safeguard and allow its actual operation to be corroborated. The gap being tested is visual text from documents was promoted to trusted system instruction.
Incorrect Answers
B:“SBOM review, sandbox network trace and blocked installer execution” would test model supply-chain files were not scanned or sandboxed before execution, not the present safeguard against visual text from documents was promoted to trusted system instruction. The expected audit evidence is “Adversarial invoice tests showing extraction without instruction execution”.
C:“Red-team adaptive prompt series and risk alert response evidence” would test adaptive jailbreak testing and abuse monitoring were inadequate, not the present safeguard against visual text from documents was promoted to trusted system instruction. The expected audit evidence is “Adversarial invoice tests showing extraction without instruction execution”.
D:“Failed hash validation on tampered model artifact in a release test” would test model artifacts themselves lack integrity controls independent of container signing, not the present safeguard against visual text from documents was promoted to trusted system instruction. The expected audit evidence is “Adversarial invoice tests showing extraction without instruction execution”.