Question 160
A user loses access to a folder, but a long-running assistant still holds an old permissions snapshot. What should protect later reads?
- Revalidate permissions at the resource-access boundary
- Assume login-time authorization lasts for every future action
- Ask the user whether access was revoked
- Keep serving cached documents until the conversation ends
Explanation
Correct Answer
A:Entitlements can change during a session. The data service should apply current authorization rather than rely indefinitely on a stale snapshot.
Incorrect Answers
B:Authentication duration does not necessarily freeze resource permissions. Later revocation must affect subsequent protected operations.
C:User self-report cannot enforce an authoritative permission change. The resource owner or policy system must determine current entitlement.
D:Session continuity does not override revocation. Cached content also needs an access policy that reflects the current authorization boundary.
Question 161
An agent checks permission to modify a record, waits several minutes, then writes after ownership has changed. What control is missing?
- A longer explanation of the earlier permission check
- A cached success result from another record
- Authorization and state preconditions at commit time
- A retry that bypasses permission checking
Explanation
Correct Answer
C:Earlier eligibility can become stale before execution. The modifying service must verify that the action remains permitted against the current record state.
Incorrect Answers
A:Narration does not update the authorization decision. The relevant issue is the state at the time the write is committed.
B:Authorization is resource-specific and time-sensitive. Success on a different record cannot establish permission for this one.
D:Skipping the check increases the risk created by the stale decision. Recovery must preserve the current authorization requirement.
Question 162
A client tool needs an API credential. Where should it normally reside in a custom application?
- In a default-value field of the model-visible tool schema
- In protected application configuration used by the executor
- Inside the tool's natural-language description
- In the conversation history so later tool calls can retrieve it
Explanation
Correct Answer
B:The application can authenticate tool execution without placing the secret in model-visible text. Access and rotation can be managed through the credential system.
Incorrect Answers
A:Schema defaults are part of the tool information exposed to the model. A credential should be injected by the authorized executor without placing its value in that context.
C:Descriptions are supplied to the model and may appear in debugging or context. They are not an appropriate secret-storage boundary.
D:Conversation persistence increases where the credential can be exposed. The executor can supply authentication without making the secret a reusable conversational fact.
Question 163
An integration needs to read calendar availability but requests permission to delete events too. Which change best aligns authorization with the task?
- Request only the permissions required for availability lookup
- Hide the deletion scope from the user interface
- Use a more capable model with the same scopes
- Keep broad permissions and promise not to use deletion
Explanation
Correct Answer
A:Unneeded authority increases consequences if the integration is misused. Permission scope should reflect the operations the application actually needs.
Incorrect Answers
B:Concealing the permission does not reduce access. It also makes the authorization less transparent to the resource owner.
C:Model capability does not change the privileges granted to the integration. Least privilege must be implemented in the authorization configuration.
D:A behavioral promise does not remove the available capability. Narrow authorization provides a stronger boundary against unintended operations.
Question 164
A browser-based integration starts an OAuth flow and later receives a callback. The flow uses a fresh state value tied to the initiating session. What should the client verify before accepting the callback?
- The authorization code is longer than a configured minimum
- The returned state matches the pending flow for that session
- The callback contains any state value, regardless of its association
- The user visited the application recently, even without a pending flow
Explanation
Correct Answer
B:State correlation helps reject unsolicited or mismatched authorization responses. It is one part of the flow and does not replace the token exchange or other required checks.
Incorrect Answers
A:Code length does not establish which session initiated authorization. The client must validate the flow association rather than infer it from appearance.
C:Presence alone does not bind the response to the initiating session. A mismatched or attacker-chosen value would still pass such a weak check.
D:A recent visit is not a binding between this response and an initiated request. The callback needs validation against the specific pending authorization state.
Question 165
A local MCP integration uses standard input/output, while a remote deployment uses HTTP. What should an architect avoid assuming?
- That tool semantics can be documented consistently
- That a local process can have security implications
- That both integrations need error handling
- That one transport's credential mechanism automatically fits the other
Explanation
Correct Answer
D:Transport and deployment context affect authentication design. Local process credentials and remote HTTP authorization require their respective supported security arrangements.
Incorrect Answers
A:Consistent capability descriptions are still useful across transports. The caution concerns credential handling, not the possibility of a clear tool contract.
B:Local execution can access files and credentials under its process permissions. Locality does not make the integration automatically harmless.
C:Both local and remote executions can fail. Error handling remains necessary even though their transport and authorization details differ.
Question 166
A response cache keys answers only by question text. Two tenants ask the same question about private financial data. What must be added to the reuse design?
- Only the generated answer's character count
- A longer expiration period for repeated questions
- Authorization scope and relevant tenant-specific input identity
- The caller's visible organization label without a stable authorization scope
Explanation
Correct Answer
C:Equivalent wording does not make private evidence equivalent. Cache reuse must preserve tenant boundaries and the data versions that determine the answer.
Incorrect Answers
A:Length cannot distinguish which tenant's records supported the answer. Two private responses can have identical sizes while containing different information.
B:Longer retention increases the time an incorrectly shared answer can be served. It does not solve the missing isolation dimension.
D:Display labels may be mutable or nonunique and do not encode effective access. A cache key must respect the actual tenant and authorization boundary.
Question 167
A vector search finds a highly relevant passage from a document the requester cannot read. What should happen before the passage enters model context?
- Remove the document title but retain its contents
- Enforce document-level access filtering
- Include the passage and ask the model not to quote it
- Use the passage only if its similarity score is high
Explanation
Correct Answer
B:Semantic relevance does not grant access. The retrieval path must apply the requester's authorization so protected text is not disclosed through the model.
Incorrect Answers
A:Removing attribution does not protect the confidential information itself. Access control must apply to the passage's content, not just its name.
C:The text has already crossed the access boundary once supplied. Output instructions do not undo unauthorized processing or possible disclosure.
D:Similarity measures relevance, not entitlement. A perfect match remains unauthorized when the requester lacks access to the document.
Question 168
Tool failures require debugging, but authorization headers contain credentials. Which logging design best supports investigation?
- Log all headers because errors are rare
- Record request metadata and redact sensitive credential fields
- Store unredacted headers in debug logs and redact only the user-facing error
- Base64-encode credential headers before writing them to logs
Explanation
Correct Answer
B:Operational identifiers and error details can support diagnosis without retaining secrets. Redaction should occur before logs enter broadly accessible systems.
Incorrect Answers
A:Low frequency does not make credential exposure acceptable. A single retained secret can be misused independently of how often failures occur.
C:The secret remains exposed in the diagnostic store even if the displayed error is safe. Redaction needs to apply before sensitive fields enter that log.
D:Base64 is a reversible representation, not secret removal. Anyone with access to those logs can recover the credential unless the sensitive field is excluded or protected appropriately.
Question 169
A Claude API request receives a rate-limit response with retry guidance. Which client behavior is appropriate?
- Respect the guidance and apply bounded backoff
- Rotate through untracked credentials to evade the limit
- Treat the response as a permanent invalid-prompt error
- Retry immediately in an unlimited tight loop
Explanation
Correct Answer
A:Rate limiting indicates that immediate repeated traffic is unlikely to help. Controlled retries should fit the request deadline and available capacity.
Incorrect Answers
B:Evading controls is not a sound capacity design. The application should manage demand and approved quotas rather than obscure its request identity.
C:A capacity limit differs from a malformed request. Classifying it correctly allows a controlled retry instead of unnecessary prompt rewriting.
D:A retry storm can worsen throttling and consume client resources. It ignores the service's explicit signal to slow requests.
Question 170
A gateway retries three times, a client retries three times, and each tool also retries. A single failure causes many requests. What should be coordinated?
- A separate full retry allowance for every nested library
- The aggregate retry count only after all layers finish retrying
- Equal per-layer retry counts without sharing attempt or deadline information
- An end-to-end retry budget across layers
Explanation
Correct Answer
D:Independent retry policies can multiply attempts. A shared deadline and attempt budget prevent recovery logic from overwhelming the same failing dependency.
Incorrect Answers
A:Independent allowances multiply across layers and can exceed the intended request budget. The design needs accounting across the full call path.
B:Observing amplification afterward helps diagnosis but does not limit it. Retry allowances need coordination before nested layers multiply attempts.
C:Equal local limits can still multiply across nested layers. A shared overall allowance is needed to bound work for the original request.