Question 330
A support agent receives a request to expose another customer's identity record. Which architecture response is required?
- Enforce object-level authorization before the record can be retrieved.
- Provide narrowly scoped credentials and enforce tenant filters in the service.
- Regard the embedded command as prompt injection from lower-trust content.
- Require an outbound destination allowlist and network isolation.
Explanation
Correct Answer
A:Preventing the data from entering model context is stronger than asking the model to conceal information it already received.
Incorrect Answers
B:Least privilege reduces the blast radius even if a malicious prompt induces an unexpected tool call; for this request, object-level identity access is the missing piece, not cross-tenant credential scope.
C:External page text is evidence, not authorization; successful defense preserves the boundary between retrieved data and trusted instructions; even if retrieved-content prompt injection succeeds, object-level identity access still needs separate verification.
D:A prompt cannot block server-side request forgery reliably; the executor must reject prohibited hosts and address ranges; the consequence is improved tool-triggered SSRF prevention, not verified object-level identity access.
Question 331
A team logs every full prompt and response in a third-party analytics system. What privacy concern must be mitigated?
- Implement an outbound destination allowlist and network isolation.
- Adopt typed allowlisted operations or isolated execution with validation instead of raw shell interpolation.
- Minimize and redact sensitive telemetry under a defined retention and access policy.
- Apply object-level authorization before the record can be retrieved.
Explanation
Correct Answer
C:Observability should not create a secondary uncontrolled store of personal or confidential content. Separate aggregated operational metrics from payload retention and apply strict access, expiry, and redaction rules to content traces.
Incorrect Answers
A:A prompt cannot block server-side request forgery reliably; the executor must reject prohibited hosts and address ranges; the effect concerns tool-triggered SSRF prevention; the success condition is prompt log data minimization.
B:Structured parameters and sandboxed privileges reduce the chance that generated strings become arbitrary commands; the prompt log data minimization criterion differs from generated shell command hardening; the gap persists.
D:Preventing the data from entering model context is stronger than asking the model to conceal information it already received; it improves object-level identity access only, leaving prompt log data minimization unproven.
Question 332
A tool accepts shell commands composed directly from model text. What control helps prevent command injection?
- Limit records using the authenticated user's data permissions.
- Keep trusted message roles assigned by the server, not asserted inside user text.
- Exercise direct jailbreaks, indirect injections, and unsafe tool-side effects separately.
- Use typed allowlisted operations or isolated execution with validation instead of raw shell interpolation.
Explanation
Correct Answer
D:Structured parameters and sandboxed privileges reduce the chance that generated strings become arbitrary commands.
Incorrect Answers
A:Prompt-level requests to avoid sensitive material do not replace backend access control at the retrieval boundary; for this request, generated shell command hardening is the missing piece, not data access filtering.
B:A user string cannot promote itself to a higher-priority protocol role; separation is essential for instruction hierarchy; this targets user role spoofing, while the system must demonstrate generated shell command hardening.
C:Different attack entry points require different boundaries; a generic refusal-rate metric cannot cover tool misuse and data leakage; for generated shell command hardening, a control centered on multi-entry threat modeling is not the appropriate substitute.
Question 333
A structured-output consumer must make safe decisions even when the API returns a successful status. Which TWO checks remain necessary? Choose TWO.
- Commit all integer values without domain checks.
- Handle refusal or incomplete-generation states before parsing.
- Validate business constraints after schema-conformant parsing.
- Consider any 2xx HTTP response a verified business transaction.
- Infer missing fields from the model’s confident wording.
Explanation
Correct Answers
B:A completed HTTP exchange may still contain a refusal or truncated response that is not a usable object. Check finish and error metadata before treating bytes as an instance of the expected JSON object.
C:Well-formed data can still contain nonexistent references or forbidden values. Use authoritative lookups and numeric range checks before the response drives an external transaction.
Incorrect Answers
A:A type restriction alone cannot prove an amount is permitted. A signed numeric field can satisfy the schema while violating payment rules or account policy.
D:Transport success is distinct from application and semantic success. A transport-level success code does not prove that the requested operation was valid or completed.
E:Language confidence does not establish data integrity. Confidence expressed in prose is not evidence that a mandatory field can be recovered correctly.
Question 334
An agent has a database credential capable of reading every tenant's data. What risk reduction is strongest?
- Issue narrowly scoped credentials and enforce tenant filters in the service.
- Retain the text inside the untrusted document channel regardless of its claimed role.
- Require object-level authorization before the record can be retrieved.
- Fix search authorization and document filtering before response generation.
Explanation
Correct Answer
A:Least privilege reduces the blast radius even if a malicious prompt induces an unexpected tool call. Use per-user or per-tenant query scopes in the database so a compromised agent credential cannot cross account boundaries.
Incorrect Answers
B:Role labels in retrieved content do not confer system authority; source provenance determines trust; the cross-tenant credential scope criterion differs from multi-criterion integration; the gap persists.
C:Preventing the data from entering model context is stronger than asking the model to conceal information it already received; this targets object-level identity access, while the system must demonstrate cross-tenant credential scope.
D:Post-generation redaction is fragile when the wrong source was already supplied; secure retrieval prevents exposure earlier; this implements retrieval authorization leak rather than the demanded cross-tenant credential scope.
Question 335
A developer wants to use model refusals as the sole protection against sending sensitive files. What is missing?
- Apply independent content and permission checks before file export.
- Implement object-level authorization before the record can be retrieved.
- Interpret the embedded command as prompt injection from lower-trust content.
- Apply typed allowlisted operations or isolated execution with validation instead of raw shell interpolation.
- Validate direct jailbreaks, indirect injections, and unsafe tool-side effects separately.
- Maintain trusted message roles assigned by the server, not asserted inside user text.
Explanation
Correct Answer
A:Safety behavior is probabilistic and cannot replace deterministic authorization or data-loss controls for external writes.
Incorrect Answers
B:Preventing the data from entering model context is stronger than asking the model to conceal information it already received; a remedy for object-level identity access is not a substitute for external data exfiltration boundary.
C:External page text is evidence, not authorization; successful defense preserves the boundary between retrieved data and trusted instructions; it changes retrieved-content prompt injection; it does not prove external data exfiltration boundary in this case.
D:Structured parameters and sandboxed privileges reduce the chance that generated strings become arbitrary commands; this may help generated shell command hardening but cannot by itself deliver external data exfiltration boundary.
E:Different attack entry points require different boundaries; a generic refusal-rate metric cannot cover tool misuse and data leakage; it improves multi-entry threat modeling only, leaving external data exfiltration boundary unproven.
F:A user string cannot promote itself to a higher-priority protocol role; separation is essential for instruction hierarchy; the scenario's controlling requirement is external data exfiltration boundary, not the related user role spoofing concern.
Question 336
A chatbot accepts pasted user instructions that imitate developer and system messages. What application behavior should remain unchanged?
- Choose typed allowlisted operations or isolated execution with validation instead of raw shell interpolation.
- Preserve trusted message roles assigned by the server, not asserted inside user text.
- Enforce independent content and permission checks before file export.
- Create narrowly scoped credentials and enforce tenant filters in the service.
- Preserve the text inside the untrusted document channel regardless of its claimed role.
Explanation
Correct Answer
B:A user string cannot promote itself to a higher-priority protocol role; separation is essential for instruction hierarchy.
Incorrect Answers
A:Structured parameters and sandboxed privileges reduce the chance that generated strings become arbitrary commands; for user role spoofing, this instead manages generated shell command hardening; the requested decision remains unresolved.
C:Safety behavior is probabilistic and cannot replace deterministic authorization or data-loss controls for external writes; because external data exfiltration boundary is its concern, user role spoofing remains outside its effect here.
D:Least privilege reduces the blast radius even if a malicious prompt induces an unexpected tool call; testing user role spoofing would still fail after improving cross-tenant credential scope.
E:Role labels in retrieved content do not confer system authority; source provenance determines trust; it changes multi-criterion integration; it does not prove user role spoofing in this case.
Question 337
An internal knowledge search index still grants access to a confidential document after the source system revokes a user’s permission. What ongoing integration control is missing?
- Use a stronger system prompt instructing Claude not to quote confidential documents.
- Increase vector-search similarity so confidential documents are retrieved only when highly relevant.
- Leave the index unchanged until the next annual document migration.
- Refresh or revalidate source document ACLs at retrieval time rather than trusting stale index permissions.
Explanation
Correct Answer
D:Indexed permissions can become outdated when document owners change access. The retrieval layer must enforce current authoritative authorization before returning content, including documents already indexed.
Incorrect Answers
A:Natural-language instructions cannot repair stale access control metadata. The model should never receive unauthorized results from the retrieval service in the first place.
B:A high relevance score is not proof of access. It can still select a forbidden document if the authorization filter relies on outdated permissions.
C:Authorization changes can take effect immediately while the index retains old records. A delayed maintenance schedule is unsuitable for revocation-sensitive access controls.
Question 338
A safety team wants to detect harmful attempts while preserving legitimate tasks. What should its threat model cover?
- Retain trusted message roles assigned by the server, not asserted inside user text.
- Introduce independent content and permission checks before file export.
- Test direct jailbreaks, indirect injections, and unsafe tool-side effects separately.
- Provide credentials from a managed secret source during deployment.
Explanation
Correct Answer
C:Different attack entry points require different boundaries; a generic refusal-rate metric cannot cover tool misuse and data leakage.
Incorrect Answers
A:A user string cannot promote itself to a higher-priority protocol role; separation is essential for instruction hierarchy; the effect concerns user role spoofing; the success condition is multi-entry threat modeling.
B:Safety behavior is probabilistic and cannot replace deterministic authorization or data-loss controls for external writes; where multi-entry threat modeling matters, a control for external data exfiltration boundary leaves that test unmet.
D:External secret storage supports rotation and access auditing without exposing keys to every code reader; where multi-entry threat modeling matters, a control for secrets at deployment boundary leaves that test unmet.
Question 339
Two transfers arrive at the same time and both appear to fit a customer’s daily limit when checked separately. What protects the limit from a race condition?
- Ask each model call to remember that another transfer might be in progress.
- Let the UI calculate the limit and trust any request reaching the backend.
- Validate each proposed amount against the original daily allowance without updating shared state.
- Check and reserve the remaining transfer allowance atomically in the authoritative transaction service.
Explanation
Correct Answer
D:A read-then-write check performed without isolation can approve two requests against the same balance. A transaction or equivalent atomic reservation enforces the cumulative limit under concurrent agent-triggered operations.
Incorrect Answers
A:Independent model turns have no reliable synchronization primitive. Natural-language reminders do not serialize database updates or prevent two simultaneous approvals.
B:Client-side calculations are bypassable and cannot safely coordinate concurrent server activity. The transaction service must own the authoritative policy decision and commit.
C:Per-request checks ignore reservations already accepted during the same interval. The rule must compare against authoritative, updated usage, including in-flight operations.
Question 340
An app uses prompt warnings but still lets Claude call a delete-account tool for any user. What is the most effective redesign?
- Mandate an appropriate domain review or deterministic eligibility check before consequential action.
- Restrict deletion to properly authorized roles with an explicit approval step.
- Enforce amount limits in the transaction service before committing the transfer.
- As an alternative, start from minimal tool permissions and expand through tested use cases.
Explanation
Correct Answer
B:Tool permissions and identity checks block forbidden actions; prompt warnings alone cannot enforce privilege. Make the delete endpoint verify the authenticated principal, required role, target account, and any explicit approval token.
Incorrect Answers
A:Generated advice is not a verified event; a separate authority must control high-impact downstream effects; testing layered authorization for destructive tools would still fail after improving human or rule-based safety gate.
C:Deterministic policy gates protect financial state independent of the model's textual self-restraint; this implements transaction limit enforcement rather than the demanded layered authorization for destructive tools.
D:Narrow capabilities limit the harm possible from unexpected outputs or malicious instructions during early operation; because minimum viable tool permissions is its concern, layered authorization for destructive tools remains outside its effect here.