ISACA Advanced in AI Audit AAIA Practice Test 36
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”.
Question 351
In a review of Threats and Vulnerabilities Specific to AI, investigators cannot establish the expected control proof: adversarial invoice tests showing extraction without instruction execution. The underlying event was that a multimodal assistant follows instructions embedded in a scanned invoice instead of extracting only invoice fields. Which control weakness is the MOST likely cause requiring focused audit work?
- Model supply-chain files were not scanned or sandboxed before execution
- Adaptive jailbreak testing and abuse monitoring were inadequate
- Model artifacts themselves lack integrity controls independent of container signing
- Visual text from documents was promoted to trusted system instruction
Explanation
Correct Answer
D:Correct. The observed condition specifically indicates visual text from documents was promoted to trusted system instruction. A meaningful audit response is to verify this diagnosis and then recommend treat ocr/image content as untrusted task data and constrain allowable actions.
Incorrect Answers
A:“Model supply-chain files were not scanned or sandboxed before execution” concerns a separate audit weakness, not the missing assurance described by visual text from documents was promoted to trusted system instruction. The correct cause is “Visual text from documents was promoted to trusted system instruction”.
B:“Adaptive jailbreak testing and abuse monitoring were inadequate” concerns a separate audit weakness, not the missing assurance described by visual text from documents was promoted to trusted system instruction. The correct cause is “Visual text from documents was promoted to trusted system instruction”.
C:“Model artifacts themselves lack integrity controls independent of container signing” concerns a separate audit weakness, not the missing assurance described by visual text from documents was promoted to trusted system instruction. The correct cause is “Visual text from documents was promoted to trusted system instruction”.
Question 352
A developer publishes a fine-tuned model endpoint with its training-set membership accessible to outsiders. 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?
- Encode untrusted model output and apply strict rendering/content security controls
- Isolate untrusted content and enforce tool-level authorization and egress checks
- Test membership inference and reduce memorization or public access as warranted
- Restrict external tool destinations and inspect sensitive content before egress
Explanation
Correct Answer
C:Correct. Test membership inference and reduce memorization or public access as warranted is aimed at the actual failure: membership-inference exposure was not included in privacy threat assessment. The audit should subsequently obtain privacy attack baseline, mitigation results and endpoint authorization tests to test execution, not merely the existence of a policy.
Incorrect Answers
A:“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 membership-inference exposure was not included in privacy threat assessment. The relevant measure is “Test membership inference and reduce memorization or public access as warranted”.
B:“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 membership-inference exposure was not included in privacy threat assessment. The relevant measure is “Test membership inference and reduce memorization or public access as warranted”.
D:“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 membership-inference exposure was not included in privacy threat assessment. The relevant measure is “Test membership inference and reduce memorization or public access as warranted”.
Question 353
In an audit of Threats and Vulnerabilities Specific to AI, managers say this measure is live: Test membership inference and reduce memorization or public access as warranted. 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?
- XSS test payload blocked by renderer and browser policy
- Privacy attack baseline, mitigation results and endpoint authorization tests
- Red-team poisoned-document results and blocked file-export calls
- Proxy block logs and tested denied tool requests containing secret strings
Explanation
Correct Answer
B:Correct. The proposed records—privacy attack baseline, mitigation results and endpoint authorization tests—are directly tied to the asserted safeguard and allow its actual operation to be corroborated. The gap being tested is membership-inference exposure was not included in privacy threat assessment.
Incorrect Answers
A:“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 membership-inference exposure was not included in privacy threat assessment. The expected audit evidence is “Privacy attack baseline, mitigation results and endpoint authorization tests”.
C:“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 membership-inference exposure was not included in privacy threat assessment. The expected audit evidence is “Privacy attack baseline, mitigation results and endpoint authorization tests”.
D:“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 membership-inference exposure was not included in privacy threat assessment. The expected audit evidence is “Privacy attack baseline, mitigation results and endpoint authorization tests”.
Question 354
In a review of Threats and Vulnerabilities Specific to AI, investigators cannot establish the expected control proof: privacy attack baseline, mitigation results and endpoint authorization tests. The underlying event was that a developer publishes a fine-tuned model endpoint with its training-set membership accessible to outsiders. Which control weakness is the MOST likely cause requiring focused audit work?
- Membership-inference exposure was not included in privacy threat assessment
- Model output is being trusted as executable presentation code
- Retrieved content is allowed to cross the instruction trust boundary
- Tool egress and data-classification checks are missing
Explanation
Correct Answer
A:Correct. The observed condition specifically indicates membership-inference exposure was not included in privacy threat assessment. A meaningful audit response is to verify this diagnosis and then recommend test membership inference and reduce memorization or public access as warranted.
Incorrect Answers
B:“Model output is being trusted as executable presentation code” concerns a separate audit weakness, not the missing assurance described by membership-inference exposure was not included in privacy threat assessment. The correct cause is “Membership-inference exposure was not included in privacy threat assessment”.
C:“Retrieved content is allowed to cross the instruction trust boundary” concerns a separate audit weakness, not the missing assurance described by membership-inference exposure was not included in privacy threat assessment. The correct cause is “Membership-inference exposure was not included in privacy threat assessment”.
D:“Tool egress and data-classification checks are missing” concerns a separate audit weakness, not the missing assurance described by membership-inference exposure was not included in privacy threat assessment. The correct cause is “Membership-inference exposure was not included in privacy threat assessment”.
Question 355
An open-source model package includes a post-install script that contacts an unknown domain. 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?
- Enforce server-side tenant authorization on every vector search
- Apply rate limits, anomaly detection and output disclosure controls
- Treat OCR/image content as untrusted task data and constrain allowable actions
- Verify provenance and execute untrusted packages in restricted build environments
Explanation
Correct Answer
D:Correct. Verify provenance and execute untrusted packages in restricted build environments is aimed at the actual failure: model supply-chain files were not scanned or sandboxed before execution. The audit should subsequently obtain sbom review, sandbox network trace and blocked installer execution to test execution, not merely the existence of a policy.
Incorrect Answers
A:“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 model supply-chain files were not scanned or sandboxed before execution. The relevant measure is “Verify provenance and execute untrusted packages in restricted build environments”.
B:“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 supply-chain files were not scanned or sandboxed before execution. The relevant measure is “Verify provenance and execute untrusted packages in restricted build environments”.
C:“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 supply-chain files were not scanned or sandboxed before execution. The relevant measure is “Verify provenance and execute untrusted packages in restricted build environments”.
Question 356
In an audit of Threats and Vulnerabilities Specific to AI, managers say this measure is live: Verify provenance and execute untrusted packages in restricted build environments. 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?
- Cross-tenant negative queries and datastore policy test results
- Extractive attack simulation, query telemetry and throttling events
- SBOM review, sandbox network trace and blocked installer execution
- Adversarial invoice tests showing extraction without instruction execution
Explanation
Correct Answer
C:Correct. The proposed records—sbom review, sandbox network trace and blocked installer execution—are directly tied to the asserted safeguard and allow its actual operation to be corroborated. The gap being tested is model supply-chain files were not scanned or sandboxed before execution.
Incorrect Answers
A:“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 model supply-chain files were not scanned or sandboxed before execution. The expected audit evidence is “SBOM review, sandbox network trace and blocked installer execution”.
B:“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 supply-chain files were not scanned or sandboxed before execution. The expected audit evidence is “SBOM review, sandbox network trace and blocked installer execution”.
D:“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 supply-chain files were not scanned or sandboxed before execution. The expected audit evidence is “SBOM review, sandbox network trace and blocked installer execution”.
Question 357
In a review of Threats and Vulnerabilities Specific to AI, investigators cannot establish the expected control proof: sbom review, sandbox network trace and blocked installer execution. The underlying event was that an open-source model package includes a post-install script that contacts an unknown domain. Which control weakness is the MOST likely cause requiring focused audit work?
- Tenant isolation is enforced only by client-provided retrieval hints
- Model supply-chain files were not scanned or sandboxed before execution
- Query-rate and extraction defenses are inadequate for model privacy
- Visual text from documents was promoted to trusted system instruction
Explanation
Correct Answer
B:Correct. The observed condition specifically indicates model supply-chain files were not scanned or sandboxed before execution. A meaningful audit response is to verify this diagnosis and then recommend verify provenance and execute untrusted packages in restricted build environments.
Incorrect Answers
A:“Tenant isolation is enforced only by client-provided retrieval hints” concerns a separate audit weakness, not the missing assurance described by model supply-chain files were not scanned or sandboxed before execution. The correct cause is “Model supply-chain files were not scanned or sandboxed before execution”.
C:“Query-rate and extraction defenses are inadequate for model privacy” concerns a separate audit weakness, not the missing assurance described by model supply-chain files were not scanned or sandboxed before execution. The correct cause is “Model supply-chain files were not scanned or sandboxed before execution”.
D:“Visual text from documents was promoted to trusted system instruction” concerns a separate audit weakness, not the missing assurance described by model supply-chain files were not scanned or sandboxed before execution. The correct cause is “Model supply-chain files were not scanned or sandboxed before execution”.
Question 358
A chatbot output is rendered as HTML by an internal dashboard and injected script executes in reviewers' browsers. 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?
- Encode untrusted model output and apply strict rendering/content security controls
- Add adversarial interaction monitoring, guardrail hardening and escalation
- Authenticate data feeds, detect anomalous submissions and quarantine suspicious updates
- Test membership inference and reduce memorization or public access as warranted
Explanation
Correct Answer
A:Correct. Encode untrusted model output and apply strict rendering/content security controls is aimed at the actual failure: model output is being trusted as executable presentation code. The audit should subsequently obtain xss test payload blocked by renderer and browser policy to test execution, not merely the existence of a policy.
Incorrect Answers
B:“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 model output is being trusted as executable presentation code. The relevant measure is “Encode untrusted model output and apply strict rendering/content security controls”.
C:“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 model output is being trusted as executable presentation code. The relevant measure is “Encode untrusted model output and apply strict rendering/content security controls”.
D:“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 model output is being trusted as executable presentation code. The relevant measure is “Encode untrusted model output and apply strict rendering/content security controls”.
Question 359
In an audit of Threats and Vulnerabilities Specific to AI, managers say this measure is live: Encode untrusted model output and apply strict rendering/content security controls. 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?
- Red-team adaptive prompt series and risk alert response evidence
- Quarantine decisions and signed provenance for accepted feedback
- Privacy attack baseline, mitigation results and endpoint authorization tests
- XSS test payload blocked by renderer and browser policy
Explanation
Correct Answer
D:Correct. The proposed records—xss test payload blocked by renderer and browser policy—are directly tied to the asserted safeguard and allow its actual operation to be corroborated. The gap being tested is model output is being trusted as executable presentation code.
Incorrect Answers
A:“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 model output is being trusted as executable presentation code. The expected audit evidence is “XSS test payload blocked by renderer and browser policy”.
B:“Quarantine decisions and signed provenance for accepted feedback” would test feedback data lacks integrity and poisoning detection, not the present safeguard against model output is being trusted as executable presentation code. The expected audit evidence is “XSS test payload blocked by renderer and browser policy”.
C:“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 model output is being trusted as executable presentation code. The expected audit evidence is “XSS test payload blocked by renderer and browser policy”.
Question 360
In a review of Threats and Vulnerabilities Specific to AI, investigators cannot establish the expected control proof: xss test payload blocked by renderer and browser policy. The underlying event was that a chatbot output is rendered as HTML by an internal dashboard and injected script executes in reviewers' browsers. Which control weakness is the MOST likely cause requiring focused audit work?
- Adaptive jailbreak testing and abuse monitoring were inadequate
- Feedback data lacks integrity and poisoning detection
- Model output is being trusted as executable presentation code
- Membership-inference exposure was not included in privacy threat assessment
Explanation
Correct Answer
C:Correct. The observed condition specifically indicates model output is being trusted as executable presentation code. A meaningful audit response is to verify this diagnosis and then recommend encode untrusted model output and apply strict rendering/content security controls.
Incorrect Answers
A:“Adaptive jailbreak testing and abuse monitoring were inadequate” concerns a separate audit weakness, not the missing assurance described by model output is being trusted as executable presentation code. The correct cause is “Model output is being trusted as executable presentation code”.
B:“Feedback data lacks integrity and poisoning detection” concerns a separate audit weakness, not the missing assurance described by model output is being trusted as executable presentation code. The correct cause is “Model output is being trusted as executable presentation code”.
D:“Membership-inference exposure was not included in privacy threat assessment” concerns a separate audit weakness, not the missing assurance described by model output is being trusted as executable presentation code. The correct cause is “Model output is being trusted as executable presentation code”.