Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 18

Question 170

An authenticated employee manipulates a document identifier and retrieves another department's confidential brief. What independently verifiable observation is needed?

  1. A production-like test shows that a user from another department cannot obtain restricted snippets through retrieval or citations.
  2. The release check confirms that response inspections show only approved fields reach model requests and audit traces.
  3. The reviewer observes that calls with missing or insufficient actor claims fail before backend execution.
  4. The release report verifies that a revoked user token cannot initiate fresh jobs or reuse a session to access protected data.

Explanation

Correct Answer

A:

Authentication proves who is calling; per-document authorization determines which evidence may be accessed; the check directly targets the condition that an employee can log into the assistant but retrieve confidential files outside their department.

Incorrect Answers

B:

Authorized retrieval can still overshare; filtering reduces unnecessary sensitive model context; this technique fits where a tool response includes confidential fields not required by the requested task, rather than the condition that an authenticated employee manipulates a document identifier and retrieves another department's confidential brief.

C:

Tool descriptions and model intentions are not an authorization system for external side effects; the trigger would be that an MCP connector accepts tool calls without checking which human authorized them, while the immediate question involves an authenticated employee manipulates a document identifier and retrieves another department's confidential brief.

D:

Long-lived delegated access outlasts valid employment and creates stale privilege exposure; the design helps when a background job retains access after an employee leaves the company, but the present constraint is that an authenticated employee manipulates a document identifier and retrieves another department's confidential brief.

Question 171

A service-account token can modify an entire customer database even though the assistant only needs status reads. What outcome would demonstrate this requirement in operation?

  1. The observed outcome is that the tool can read allowed orders but cannot change payments or account roles.
  2. The regression suite verifies that calls with missing or insufficient actor claims fail before backend execution.
  3. A production-like test shows that response inspections show only approved fields reach model requests and audit traces.
  4. The release check confirms that changing the submitted tenant ID does not reveal another organization's records.

Explanation

Correct Answer

A:

Least-privilege credentials limit compromise impact and avoid unnecessary administrative capability; an acceptance test needs to cover cases where a service stores a broad administrator token so a tool can read order statuses.

Incorrect Answers

B:

Tool descriptions and model intentions are not an authorization system for external side effects; this applies when an MCP connector accepts tool calls without checking which human authorized them, not this situation where a service-account token can modify an entire customer database even though the assistant only needs status reads.

C:

Authorized retrieval can still overshare; filtering reduces unnecessary sensitive model context; the alternative helps if a tool response includes confidential fields not required by the requested task, whereas this scenario involves a service-account token can modify an entire customer database even though the assistant only needs status reads.

D:

Untrusted request parameters cannot establish data-access authority across customers; it is justified if a summarization tool trusts client-provided tenant IDs without checking the authenticated session, but here a service-account token can modify an entire customer database even though the assistant only needs status reads.

Question 172

A tenant-isolation test succeeds for ordinary clients but fails when a caller changes the tenant parameter. Which observed test outcome would justify sign-off?

  1. The scenario test demonstrates that changing the submitted tenant ID does not reveal another organization's records.
  2. The end-to-end check confirms that calls with missing or insufficient actor claims fail before backend execution.
  3. The observed outcome is that a revoked user token cannot initiate fresh jobs or reuse a session to access protected data.
  4. The regression suite verifies that prompt-injection tests cannot retrieve stored secrets because the backend denies the requests.

Explanation

Correct Answer

A:

Untrusted request parameters cannot establish data-access authority across customers; the control should hold even if a summarization tool trusts client-provided tenant IDs without checking the authenticated session.

Incorrect Answers

B:

Tool descriptions and model intentions are not an authorization system for external side effects; the other case is when an MCP connector accepts tool calls without checking which human authorized them, but the actual risk is that a tenant-isolation test succeeds for ordinary clients but fails when a caller changes the tenant parameter.

C:

Long-lived delegated access outlasts valid employment and creates stale privilege exposure; this makes sense when a background job retains access after an employee leaves the company, but the requested boundary concerns a tenant-isolation test succeeds for ordinary clients but fails when a caller changes the tenant parameter.

D:

Prompts guide behavior but do not guarantee prevention of unauthorized disclosure or tool access; this technique fits where a product team treats a prompt instruction to 'never disclose secrets' as the sole data boundary, not the present case where a tenant-isolation test succeeds for ordinary clients but fails when a caller changes the tenant parameter.

Question 173

A connector receives a valid tool request but has no record tying it to the permissions of the original user. Which sign-off finding best checks the actual boundary?

  1. A controlled trial finds that a revoked user token cannot initiate fresh jobs or reuse a session to access protected data.
  2. The measured evidence confirms that prompt-injection tests cannot retrieve stored secrets because the backend denies the requests.
  3. The scenario test demonstrates that calls with missing or insufficient actor claims fail before backend execution.
  4. The end-to-end check confirms that response inspections show only approved fields reach model requests and audit traces.
  5. The observed outcome is that a user from another department cannot obtain restricted snippets through retrieval or citations.

Explanation

Correct Answer

C:

Tool descriptions and model intentions are not an authorization system for external side effects; the safeguard must work when an MCP connector accepts tool calls without checking which human authorized them.

Incorrect Answers

A:

Long-lived delegated access outlasts valid employment and creates stale privilege exposure; it serves a setting where a background job retains access after an employee leaves the company, rather than the condition that a connector receives a valid tool request but has no record tying it to the permissions of the original user.

B:

Prompts guide behavior but do not guarantee prevention of unauthorized disclosure or tool access; that is defensible if a product team treats a prompt instruction to 'never disclose secrets' as the sole data boundary, while the immediate question involves a connector receives a valid tool request but has no record tying it to the permissions of the original user.

D:

Authorized retrieval can still overshare; filtering reduces unnecessary sensitive model context; this applies when a tool response includes confidential fields not required by the requested task, whereas the current issue is that a connector receives a valid tool request but has no record tying it to the permissions of the original user.

E:

Authentication proves who is calling; per-document authorization determines which evidence may be accessed; the alternative helps if an employee can log into the assistant but retrieve confidential files outside their department, whereas the release criterion addresses a connector receives a valid tool request but has no record tying it to the permissions of the original user.

Question 174

A departing analyst's background tasks continue pulling files days after their access was revoked. What validation evidence would address this failure mode?

  1. The representative case test finds that the tool can read allowed orders but cannot change payments or account roles.
  2. The negative-case evaluation confirms that calls with missing or insufficient actor claims fail before backend execution.
  3. The audit sample demonstrates that a revoked user token cannot initiate fresh jobs or reuse a session to access protected data.
  4. The measured evidence confirms that response inspections show only approved fields reach model requests and audit traces.

Explanation

Correct Answer

C:

Long-lived delegated access outlasts valid employment and creates stale privilege exposure; the test covers the original risk that a background job retains access after an employee leaves the company.

Incorrect Answers

A:

Least-privilege credentials limit compromise impact and avoid unnecessary administrative capability; that belongs to situations where a service stores a broad administrator token so a tool can read order statuses, not this situation where a departing analyst's background tasks continue pulling files days after their access was revoked.

B:

Tool descriptions and model intentions are not an authorization system for external side effects; it is relevant if an MCP connector accepts tool calls without checking which human authorized them, whereas this scenario involves a departing analyst's background tasks continue pulling files days after their access was revoked.

D:

Authorized retrieval can still overshare; filtering reduces unnecessary sensitive model context; the other case is when a tool response includes confidential fields not required by the requested task, instead of when a departing analyst's background tasks continue pulling files days after their access was revoked.

Question 175

A support agent can reset passwords despite only needing to search troubleshooting articles. Which action addresses the root concern without unnecessary scope?

  1. For this workflow, remove unused password-reset tooling from the agent's available capabilities.
  2. For the affected service, disable unneeded actions immediately and use logging as additional detection.
  3. In the project scope, partition tool sets by task role and grant capability only when relevant to the current workflow.
  4. Within the rollout design, review invocation frequency, misuse paths, ownership and necessity before retaining the connector.

Explanation

Correct Answer

A:

Capability removal prevents risky actions at their source instead of depending on model restraint; a representative trial should prove that attempted password resets fail because no such executable tool exists for the agent.

Incorrect Answers

B:

Observability records misuse but cannot itself prevent actions made possible by excessive permissions; that would be useful if a pilot shows unexpected tool invocations but its team proposes logging only, but the requested boundary concerns a support agent can reset passwords despite only needing to search troubleshooting articles.

C:

Task-scoped exposure limits attack surface and cuts tool-choice confusion in large deployments; it serves a setting where an agent's tool catalog grows as every new team adds a connector to a common pool, not the present case where a support agent can reset passwords despite only needing to search troubleshooting articles.

D:

Tool governance needs evidence of a real task requirement and proportionate privilege; that is defensible if an architect wants to measure whether a newly added connector increases risk more than value, while this problem involves a support agent can reset passwords despite only needing to search troubleshooting articles.

Question 176

A research agent has twenty near-identical tools and frequently selects the wrong one. What is the most appropriate initial architecture decision?

  1. At the system boundary, enforce server-side refund bounds, approvals and idempotency within the tool implementation.
  2. For the primary workstream, consolidate overlapping tools behind distinct, task-centered interfaces and explicit descriptions.
  3. For this workflow, partition tool sets by task role and grant capability only when relevant to the current workflow.
  4. For the affected service, review invocation frequency, misuse paths, ownership and necessity before retaining the connector.

Explanation

Correct Answer

B:

Reducing ambiguous selection improves tool reliability and limits unnecessary decision branches; the implementation should produce evidence that tool-selection tests favor the correct interface across similar natural-language tasks.

Incorrect Answers

A:

Tool schemas alone do not prevent malicious or erroneous high-impact values without backend validation; it targets cases where an automation needs a refund action but the tool accepts arbitrary monetary values, while the immediate question involves a research agent has twenty near-identical tools and frequently selects the wrong one.

C:

Task-scoped exposure limits attack surface and cuts tool-choice confusion in large deployments; that belongs to situations where an agent's tool catalog grows as every new team adds a connector to a common pool, whereas the current issue is that a research agent has twenty near-identical tools and frequently selects the wrong one.

D:

Tool governance needs evidence of a real task requirement and proportionate privilege; it is relevant if an architect wants to measure whether a newly added connector increases risk more than value, whereas the release criterion addresses a research agent has twenty near-identical tools and frequently selects the wrong one.

Question 177

A purchasing assistant is allowed to change an order status and issue refunds only within signed policy limits. Its current connector has broad database writes and accepts unconstrained refund amounts. Which TWO tool-boundary changes would enforce the permitted actions independently of prompt compliance? Choose TWO.

  1. For the intended users, expose a validated status-update operation with server-side field and transition checks.
  2. Within the delivery plan, consolidate overlapping tools behind distinct, task-centered interfaces and explicit descriptions.
  3. At the system boundary, review invocation frequency, misuse paths, ownership and necessity before retaining the connector.
  4. For the primary workstream, enforce server-side refund bounds, approvals and idempotency within the tool implementation.
  5. In the proposed design, partition tool sets by task role and grant capability only when relevant to the current workflow.

Explanation

Correct Answers

A:

Narrow application endpoints allow the required mutation without granting general table write authority; acceptance evidence must confirm that tests reject updates to price, recipient or invalid order states while allowing approved transitions; another obligation arises when an automation needs a refund action but the tool accepts arbitrary monetary values.

D:

Tool schemas alone do not prevent malicious or erroneous high-impact values without backend validation; one operational check should confirm that out-of-policy amounts are rejected regardless of Claude's argument text; another obligation arises when a purchasing assistant requests unrestricted database write access to update one order status field.

Incorrect Answers

B:

Reducing ambiguous selection improves tool reliability and limits unnecessary decision branches; that fits cases where a research agent has twenty near-identical tools and frequently selects the wrong one, but here a purchasing assistant requests unrestricted database write access to update one order status field; another obligation arises when an automation needs a refund action but the tool accepts arbitrary monetary values.

C:

Tool governance needs evidence of a real task requirement and proportionate privilege; this control is needed when an architect wants to measure whether a newly added connector increases risk more than value, instead of when an automation needs a refund action but the tool accepts arbitrary monetary values; another obligation arises when a purchasing assistant requests unrestricted database write access to update one order status field.

E:

Task-scoped exposure limits attack surface and cuts tool-choice confusion in large deployments; it serves a setting where an agent's tool catalog grows as every new team adds a connector to a common pool, whereas the problem presented is that an automation needs a refund action but the tool accepts arbitrary monetary values; another obligation arises when a purchasing assistant requests unrestricted database write access to update one order status field.

Question 178

A pilot shows unexpected tool invocations but its team proposes logging only. What should be incorporated into the approved design?

  1. Under the reviewed approach, disable unneeded actions immediately and use logging as additional detection.
  2. For this production path, consolidate overlapping tools behind distinct, task-centered interfaces and explicit descriptions.
  3. For the intended users, partition tool sets by task role and grant capability only when relevant to the current workflow.
  4. Within the delivery plan, remove unused password-reset tooling from the agent's available capabilities.

Explanation

Correct Answer

A:

Observability records misuse but cannot itself prevent actions made possible by excessive permissions; a clear sign-off signal is that adversarial prompts cannot execute retired tools and remaining calls retain audit traces.

Incorrect Answers

B:

Reducing ambiguous selection improves tool reliability and limits unnecessary decision branches; its scope includes cases where a research agent has twenty near-identical tools and frequently selects the wrong one, not the present case where a pilot shows unexpected tool invocations but its team proposes logging only.

C:

Task-scoped exposure limits attack surface and cuts tool-choice confusion in large deployments; it targets cases where an agent's tool catalog grows as every new team adds a connector to a common pool, while this problem involves a pilot shows unexpected tool invocations but its team proposes logging only.

D:

Capability removal prevents risky actions at their source instead of depending on model restraint; the approach helps when a support agent can reset passwords despite only needing to search troubleshooting articles, and not where a pilot shows unexpected tool invocations but its team proposes logging only.

Question 179

An agent's tool catalog grows as every new team adds a connector to a common pool. Which approach should the design team choose?

  1. In the initial solution, remove unused password-reset tooling from the agent's available capabilities.
  2. In the proposed design, consolidate overlapping tools behind distinct, task-centered interfaces and explicit descriptions.
  3. Under the reviewed approach, partition tool sets by task role and grant capability only when relevant to the current workflow.
  4. For this production path, expose a validated status-update operation with server-side field and transition checks.

Explanation

Correct Answer

C:

Task-scoped exposure limits attack surface and cuts tool-choice confusion in large deployments; one measurable outcome is that unrelated departments' connectors are absent from each worker's visible inventory.

Incorrect Answers

A:

Capability removal prevents risky actions at their source instead of depending on model restraint; this technique fits where a support agent can reset passwords despite only needing to search troubleshooting articles, but the present constraint is that an agent's tool catalog grows as every new team adds a connector to a common pool.

B:

Reducing ambiguous selection improves tool reliability and limits unnecessary decision branches; the trigger would be that a research agent has twenty near-identical tools and frequently selects the wrong one, whereas the current issue is that an agent's tool catalog grows as every new team adds a connector to a common pool.

D:

Narrow application endpoints allow the required mutation without granting general table write authority; that fits cases where a purchasing assistant requests unrestricted database write access to update one order status field, while the required safeguard concerns an agent's tool catalog grows as every new team adds a connector to a common pool.

Question 180

An automation needs a refund action but the tool accepts arbitrary monetary values. Where should the solution architect focus the implementation?

  1. At this decision point, enforce server-side refund bounds, approvals and idempotency within the tool implementation.
  2. As a candidate control, partition tool sets by task role and grant capability only when relevant to the current workflow.
  3. For this deployment, remove unused password-reset tooling from the agent's available capabilities.
  4. In the proposed design, expose a validated status-update operation with server-side field and transition checks.

Explanation

Correct Answer

A:

Tool schemas alone do not prevent malicious or erroneous high-impact values without backend validation; a suitable release result is that out-of-policy amounts are rejected regardless of Claude's argument text.

Incorrect Answers

B:

Task-scoped exposure limits attack surface and cuts tool-choice confusion in large deployments; the alternative helps if an agent's tool catalog grows as every new team adds a connector to a common pool, instead of when an automation needs a refund action but the tool accepts arbitrary monetary values.

C:

Capability removal prevents risky actions at their source instead of depending on model restraint; it is justified if a support agent can reset passwords despite only needing to search troubleshooting articles, not the case in which an automation needs a refund action but the tool accepts arbitrary monetary values.

D:

Narrow application endpoints allow the required mutation without granting general table write authority; its scope includes cases where a purchasing assistant requests unrestricted database write access to update one order status field, whereas the problem presented is that an automation needs a refund action but the tool accepts arbitrary monetary values.

Claude Certified Architect - Professional (CCAR-P) Practice Tests