View Full Anthropic CCA-F Exam Dumps and Practice Test Dumps.

 

Question 361

An agent needs to access customer records from a database. Which approach best follows the principle of least privilege?

1. Give the agent administrator access to the entire database
2. Allow access only to the required records and operations
3. Give the agent unrestricted write permissions
4. Allow the model to create new database permissions

Correct Answer: 2

Explanation

Least privilege means providing only the permissions necessary to perform the intended task. If an agent only needs to retrieve customer records, it should not receive unrestricted database administration privileges or permission to modify unrelated tables. Access should be scoped to the required resources and operations, ideally through narrowly designed tools or service accounts. Authorization should also be enforced independently of the model because a model may make mistakes or encounter malicious instructions. Restricting permissions reduces the potential impact of accidental tool use, compromised credentials, prompt injection, or incorrect reasoning.

Question 362

A user asks an agent to book a hotel room but does not provide the destination. What should the agent do?

1. Choose the user’s previous destination
2. Book a room in the nearest city
3. Ask the user to specify the destination
4. Select the cheapest available hotel

Correct Answer: 3

Explanation

The destination is a required piece of information for completing a hotel booking. Guessing the destination based on previous activity or choosing an arbitrary location could result in a booking that does not meet the user’s needs. The agent should ask a concise clarification question before performing the external action. Once the destination is known, additional information such as dates, number of guests, room type, and budget may also be required. This illustrates the importance of resolving critical ambiguity before executing actions that create financial or logistical commitments.

Question 363

Which practice can reduce the risk of an agent exposing sensitive information through logs?

1. Logging every model message in full
2. Sending all tool responses to public analytics systems
3. Redacting or excluding unnecessary sensitive fields
4. Disabling authentication for the logging system

Correct Answer: 3

Explanation

Sensitive information should be minimized in logs whenever it is not required for operational or security purposes. Applications can redact fields such as passwords, payment information, authentication tokens, or unnecessary personal data before writing logs. Access to the logs should also be restricted, and retention periods should be appropriate to the organization’s requirements. Logging everything may appear useful for debugging, but it increases the consequences of a log compromise. A balanced logging strategy captures enough information to diagnose failures and audit important actions without unnecessarily duplicating sensitive customer information.

Question 364

An agent receives a request containing a valid customer ID but the requester is not authorized to access that customer. What should happen?

1. The request should be denied by the authorization layer
2. The agent should return the information because the ID is valid
3. The agent should ask the user to guess another ID
4. The agent should bypass the authorization check

Correct Answer: 1

Explanation

A valid identifier does not establish that the requester has permission to access the associated resource. Authorization must determine whether the authenticated requester can perform the requested operation on that specific customer record. The authorization check should ideally occur in the application or service layer rather than relying only on the model. This prevents the agent from accidentally exposing information when a user supplies a valid but unauthorized identifier. Resource-level authorization is particularly important in multi-tenant systems where different users and organizations may have access to different subsets of data.

Question 365

An agent is processing a request that may take several minutes. Which design can help prevent the workflow from running indefinitely?

1. Remove all execution limits
2. Use a timeout and defined failure or continuation state
3. Retry continuously without delay
4. Allow the model to reset the timeout

Correct Answer: 2

Explanation

Long-running workflows should have explicit timeout controls so that a temporary problem cannot cause indefinite execution. When a timeout occurs, the application should transition the workflow into a defined state such as timed_out, failed, or pending_retry. If the operation can safely resume, the system can provide a controlled continuation mechanism. Unlimited execution can increase costs and consume resources while providing no guarantee of completion. Timeouts should be implemented at the application or orchestration layer rather than depending on the model to recognize that it has been running too long.

Question 366

Why is provenance metadata useful when an agent provides information retrieved from external sources?

1. It identifies where the information came from and can support verification
2. It guarantees that the information is always correct
3. It prevents all prompt-injection attacks
4. It eliminates the need for source validation

Correct Answer: 1

Explanation

Provenance metadata can identify the source, timestamp, document, or retrieval context associated with information presented to the agent. This allows users and applications to understand where a claim originated and can help identify stale, conflicting, or questionable information. Provenance does not automatically guarantee accuracy, because the source itself may contain errors. However, it provides useful context for verification and auditing. In retrieval-based systems, maintaining source identifiers and timestamps can also help the application select fresher information when multiple sources provide conflicting results.

Question 367

An agent is asked to update a customer’s email address, but the new address fails validation. What should the system do?

1. Save the invalid address anyway
2. Replace it with a guessed address
3. Reject the update and request a valid address
4. Delete the customer’s account

Correct Answer: 3

Explanation

Input validation should occur before a persistent customer record is modified. If the new email address does not satisfy the application’s validation rules, the update should be rejected and the user should be asked to provide a valid value. Guessing an address could cause communications to be sent to the wrong recipient. Validation can include syntax checks and, where appropriate, verification workflows such as confirmation emails. The agent should accurately explain that the requested update could not be applied rather than claiming success. Application-level validation should remain active regardless of the model’s generated response.

Question 368

Which situation most clearly requires the agent to request clarification before acting?

1. The user asks for a summary of a named document
2. The user asks to permanently delete “the account” when multiple accounts exist
3. The user asks for today’s date
4. The user asks to calculate a known percentage

Correct Answer: 2

Explanation

Permanent deletion is a destructive action, and the phrase “the account” is ambiguous when multiple accounts exist. The agent should identify exactly which account the user intends before performing the deletion. This combines two important safety considerations: ambiguity and irreversibility. Selecting an account based on an unsupported assumption could cause serious data loss. The system should also verify authorization and, where appropriate, obtain explicit confirmation immediately before deletion. Low-risk informational requests generally require less clarification when their intended target and expected output are already clear.

Question 369

What is the main purpose of an evaluation dataset for an agent application?

1. To replace production monitoring completely
2. To provide repeatable scenarios for measuring system behavior
3. To store user passwords
4. To give the model unlimited tool permissions

Correct Answer: 2

Explanation

An evaluation dataset provides repeatable test cases that can be used to measure an agent’s behavior over time. It may include common user requests, edge cases, tool-selection scenarios, security tests, and previously observed failures. Developers can compare results across prompt, model, tool, or application changes to detect regressions. Evaluation data does not replace production monitoring because real-world usage can expose situations that were not represented in the test set. Sensitive production information should also be handled carefully when constructing evaluation datasets, with appropriate anonymization or synthetic data where possible.

Question 370

An agent uses a third-party service to process customer information. Which principle should guide the data sent to that service?

1. Send all available customer information for maximum context
2. Send only the information necessary for the requested operation
3. Send confidential credentials with every request
4. Send unrelated customer records to improve accuracy

Correct Answer: 2

Explanation

Data minimization should also apply when information is sent to third-party services. The application should provide only the fields required to perform the requested operation rather than transmitting an entire customer record. This reduces privacy exposure and limits the consequences of a third-party compromise or accidental disclosure. Sensitive credentials should never be included merely because they are available. Depending on the service and data involved, organizations may also need contractual, regulatory, or security controls governing third-party processing. Minimizing data is a practical way to reduce unnecessary exposure while still supporting the required workflow.

Question 371

An agent is configured to call a payment tool whenever the user mentions “pay.” Why is this design risky?

1. The keyword may not indicate that the user actually wants a payment executed
2. Payment tools never require authorization
3. The model cannot process financial requests
4. Keywords automatically create valid transactions

Correct Answer: 1

Explanation

A keyword alone does not establish the user’s intent to execute a financial transaction. A user might mention “pay” while asking a general question, checking a previous payment, or discussing an unrelated issue. Triggering a payment tool based solely on a keyword can therefore create unintended financial side effects. The agent should understand the requested action, identify the target and amount, validate authorization, and obtain confirmation where required. Tool invocation should be based on structured intent and validated parameters rather than simplistic keyword matching.

Question 372

What should happen if a tool requires an integer quantity, but the agent provides a text value such as “many”?

1. Convert it to an arbitrary number
2. Accept it without validation
3. Reject the invalid parameter and request a valid quantity
4. Use the maximum allowed quantity

Correct Answer: 3

Explanation

Tool parameters should conform to their defined schemas. If quantity requires an integer and the supplied value is “many,” the application should reject the request rather than guessing a numeric interpretation. The user or agent can then provide a valid quantity. Strict parameter validation prevents malformed data from reaching downstream systems and makes tool behavior predictable. For operations involving purchases, inventory, or other side effects, guessing a quantity could have financial or operational consequences. Validation should occur before execution and should produce a structured error that can be handled appropriately.

Question 373

An agent’s tool response contains a timestamp showing that the data is several days old. The user asks for the current status. What should the agent consider?

1. Whether the stale data is still appropriate or whether fresh data should be retrieved
2. Treat the old information as current automatically
3. Delete the timestamp
4. Change the timestamp to today’s date

Correct Answer: 1

Explanation

Freshness matters when the user asks for current information. A timestamp showing that data is several days old may indicate that the result is no longer reliable for the requested purpose. The agent should consider the application’s freshness requirements and, when necessary, retrieve a newer result from the authoritative source. It should not modify or conceal the timestamp. Freshness policies can vary by domain: inventory may require near-real-time information, while historical records may remain valid for much longer. Explicit timestamps help applications and users make informed decisions about data reliability.

Question 374

Which approach is most appropriate for handling a tool that performs an irreversible operation?

1. Give the tool unrestricted access to every user
2. Require appropriate authorization and stronger safeguards before execution
3. Automatically execute whenever the model mentions the operation
4. Remove all audit logging

Correct Answer: 2

Explanation

Irreversible operations deserve stronger controls because mistakes may be difficult or impossible to undo. Appropriate safeguards can include narrow permissions, precise target validation, explicit confirmation, human approval for high-impact actions, audit logging, and execution limits. The exact controls should reflect the operation’s potential consequences. The model should not be able to bypass these protections simply by generating a tool call. Separating destructive capabilities from read-only tools can also make permissions easier to manage. These measures help ensure that an agent’s autonomy remains proportional to the risk associated with the requested operation.

Question 375

A model update changes how an agent selects tools, causing several previously successful workflows to fail. What should the development team do?

1. Ignore the failures because the model is newer
2. Remove the affected workflows
3. Use regression testing to compare behavior and investigate the changes
4. Give the model additional unrestricted permissions

Correct Answer: 3

Explanation

A model update can change tool selection, reasoning patterns, output formats, and other behaviors even when the surrounding application has not changed. Regression testing provides a systematic way to compare the new model against previously validated workflows. The team can identify which scenarios changed, determine whether the differences are acceptable, and modify prompts, tools, validation, or model configuration as necessary. Production monitoring can provide additional evidence after deployment. Versioning the model, prompts, and tools makes it easier to reproduce and investigate behavioral changes.

Question 376

An agent receives a request to modify a user’s account but the user’s identity has not been authenticated. What should happen?

1. Perform the modification because the user knows the account details
2. Ask the model to determine the user’s identity
3. Require appropriate authentication before proceeding
4. Modify every account matching the user’s name

Correct Answer: 3

Explanation

Sensitive account modifications require reliable identification of the requester. Knowing an account number or other public information does not necessarily prove that the person is authorized to make changes. The application should require the appropriate authentication mechanism before allowing the agent to perform the operation. Authentication should be followed by authorization to determine whether that authenticated identity has permission to modify the specific account. These controls should be implemented outside the model so that a conversational instruction cannot bypass the application’s identity and access requirements.

Question 377

Which design best prevents an agent from accidentally using a production database during testing?

1. Use separate environments and credentials for testing and production
2. Give developers production administrator credentials
3. Use production data for every test
4. Allow the model to choose the database automatically

Correct Answer: 1

Explanation

Separate environments and credentials create an important boundary between testing and production. Development and testing workflows should normally use dedicated resources and appropriately restricted accounts, while production credentials should be protected from routine experimentation. This reduces the risk that an incorrect test, malformed request, or agent mistake modifies real customer data. Configuration should explicitly identify the target environment rather than relying on the model to select it. Additional safeguards such as network restrictions, permission scopes, deployment approvals, and monitoring can further reduce the chance of accidental production access.

Question 378

A user asks an agent to cancel a subscription immediately, but the system requires a confirmation step before cancellation. What should the agent do?

1. Bypass the confirmation requirement
2. Follow the required confirmation workflow
3. Cancel the subscription silently
4. Disable the cancellation policy

Correct Answer: 2

Explanation

Application policies governing consequential actions should not be bypassed merely because the user requests immediate execution. If cancellation requires confirmation, the agent should follow the established workflow and obtain the required confirmation before executing the operation. This ensures that the user understands the action and gives the application an opportunity to enforce additional authorization or business rules. The confirmation requirement should be implemented at the application or tool layer rather than relying entirely on the model. If the user does not complete the required confirmation, the cancellation should remain unexecuted.

Question 379

What is an important benefit of using structured status values such as pending, completed, and failed in an agent workflow?

1. They make workflow state explicit and easier for software to manage
2. They guarantee successful execution
3. They eliminate the need for error handling
4. They allow the model to ignore incomplete operations

Correct Answer: 1

Explanation

Explicit status values provide a consistent representation of workflow progress. Software can use states such as pending, in_progress, completed, failed, and timed_out to determine what actions are appropriate next. This is especially useful for multi-step workflows where some operations may succeed while others fail. Structured states also improve monitoring, reporting, and recovery because the application does not have to infer completion from free-form text. They do not guarantee that an operation succeeds, but they make the actual state easier to track and communicate accurately.

Question 380

An agent has completed the requested operation successfully. What should it avoid doing afterward?

1. Recording the appropriate completion status
2. Reporting the successful result accurately
3. Repeating the same side-effecting operation unnecessarily
4. Returning relevant information to the user

Correct Answer: 3

Explanation

Once a side-effecting operation has successfully completed, repeating it without a valid reason can create duplicate or unintended effects. For example, submitting the same payment, sending the same email, or creating the same appointment twice can cause significant problems. The workflow should record a completed state and use that state to prevent unnecessary repetition. Idempotency mechanisms can provide additional protection when retries are unavoidable. Accurate completion reporting also helps the user understand what happened. Clear termination conditions are therefore an important part of reliable agent orchestration.