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

 

Question 281

An agent needs to update a customer’s phone number. Which approach best limits unintended changes?

1. Give the agent access to the entire customer database
2. Allow modification of every customer field
3. Use a narrowly scoped tool that updates only the phone number
4. Allow the model to execute arbitrary database queries

Correct Answer: 3

Explanation

A narrowly scoped tool limits the agent’s ability to modify unrelated information. If the requested task is to change a phone number, the tool should ideally accept the customer identifier and the new phone number while preventing changes to unrelated fields. This follows least-privilege principles and makes the operation easier to validate and audit. Broad database access creates unnecessary risk because an incorrect model decision could affect other customer information. Backend authorization should also verify that the requesting user is allowed to modify the specific customer’s record before the update is performed.

Question 282

A user asks an agent to send money to “John,” but several recipients named John are available. What should the agent do?

1. Send the money to the first matching recipient
2. Ask the user to identify the intended recipient
3. Send money to all matching recipients
4. Select the recipient with the largest transaction history

Correct Answer: 2

Explanation

A financial transfer should not be executed when the intended recipient is ambiguous. If multiple recipients match the name “John,” the agent should ask the user for additional identifying information, such as an account identifier or the recipient’s full details. Choosing the first result or using transaction history as a selection rule would introduce an unsupported assumption. Sending money to every matching recipient could cause serious financial loss. Because the action has an external and potentially irreversible effect, resolving the ambiguity before execution is an important safety measure.

Question 283

Which practice can help prevent sensitive information from appearing in debugging logs?

1. Log all request and response data without filtering
2. Disable every form of application monitoring
3. Store credentials directly in log messages
4. Apply redaction and data-minimization rules to logs

Correct Answer: 4

Explanation

Logs should contain enough information to support troubleshooting and auditing without unnecessarily exposing sensitive data. Redaction can remove passwords, authentication tokens, payment information, personal identifiers, or other sensitive values before records are stored. Data minimization further ensures that information unrelated to debugging is not captured in the first place. Completely disabling logging can make investigations difficult, while storing credentials in logs creates an additional security vulnerability. A well-designed logging strategy balances observability with privacy and security requirements, using controlled access and appropriate retention periods for the information that is actually necessary.

Question 284

An agent is given a tool that can both read and delete files. The user only requests a file listing. Which capability should be used?

1. The read-only listing operation
2. The delete operation
3. Both operations
4. A general-purpose system administration command

Correct Answer: 1

Explanation

The agent should use the least powerful capability that can satisfy the user’s request. A file listing requires only read access, so there is no reason to invoke a deletion capability or a broader administrative command. Limiting operations to the minimum necessary reduces the risk of unintended side effects and follows the principle of least privilege. Clear tool boundaries also make it easier to audit which capabilities were used. Even when a tool technically provides multiple operations, the application should ensure that the agent cannot unnecessarily perform destructive actions when a simple read operation is sufficient.

Question 285

Why should an AI agent maintain clear boundaries between trusted instructions and external content?

1. External content can contain instructions intended to manipulate agent behavior
2. External content is always accurate
3. Trusted instructions are unnecessary when using tools
4. It allows external webpages to change authorization policies

Correct Answer: 1

Explanation

External content such as webpages, emails, uploaded documents, and retrieved text may contain instructions that are not authorized by the application. These instructions can be designed to manipulate the model into revealing information, ignoring policies, or invoking tools inappropriately. Maintaining a clear boundary between trusted instructions and untrusted content helps preserve the intended instruction hierarchy. Authorization should also be enforced independently by the application. External content can still provide useful information, but it should not automatically gain the authority to change system behavior or security permissions.

Question 286

An agent’s tool returns a structured response showing that a request is pending. What should the agent communicate?

1. That the request definitely failed
2. That the request is still pending rather than completed
3. That the request completed successfully
4. That the tool response should be ignored

Correct Answer: 2

Explanation

A pending status means the operation has not reached a final state. The agent should communicate this accurately rather than claiming success or failure without evidence. Depending on the workflow, the application may provide a status identifier, allow polling, or notify the user when the operation completes. Treating a pending operation as successful could cause the user to believe that an action has already taken effect. Structured workflow states help agents and applications distinguish between completed, failed, and in-progress operations and determine the appropriate next step.

Question 287

What is a key reason to use bounded retries in an agent workflow?

1. To make the agent retry every failure forever
2. To avoid consuming unlimited resources during repeated failures
3. To eliminate all external service errors
4. To guarantee that every request succeeds

Correct Answer: 2

Explanation

Bounded retries place a clear limit on how many times an operation can be attempted. This prevents an agent from entering an endless retry loop when a service remains unavailable or an error cannot be resolved automatically. Retry policies can also use backoff to reduce pressure on external services. Bounded retries do not guarantee success and should not be applied blindly to operations that may create duplicate side effects. For such operations, idempotency or transaction protection may be required. Once the retry limit is reached, the system should report the unresolved condition accurately.

Question 288

An agent is retrieving private documents for a user. What should determine whether each document can be returned?

1. Whether the document has a short title
2. Whether the model finds the document interesting
3. Whether the authenticated user is authorized to access the document
4. Whether another customer accessed it previously

Correct Answer: 3

Explanation

Access to private documents should be controlled by authorization enforced by the application or retrieval system. The model should not independently decide whether a document is safe to return. The system should verify the authenticated user’s permissions against the requested resource before the document enters the agent’s context. This is particularly important in multi-tenant environments where documents may belong to different customers or teams. Authorization filtering at the retrieval layer helps prevent unauthorized information from reaching the model in the first place and provides a stronger security boundary than relying solely on natural-language instructions.

Question 289

An agent receives an invalid JSON response from a tool that is expected to return structured data. What should happen?

1. Validate the response and handle the parsing failure according to the error policy
2. Assume the JSON is correct
3. Execute the next high-impact action immediately
4. Convert arbitrary text into a successful status

Correct Answer: 1

Explanation

If a tool is expected to return structured JSON but instead produces invalid data, the application should detect the parsing or schema-validation failure before using the response downstream. Depending on the workflow, it may retry if the error is temporary, report the failure, or invoke an alternative recovery path. Executing a consequential operation using malformed data can produce incorrect results. Structured output contracts are useful because they allow applications to detect unexpected responses automatically. Validation should therefore occur before downstream processing, especially when the result could trigger financial, administrative, or destructive actions.

Question 290

Which approach best supports reproducibility when investigating an agent failure?

1. Delete previous execution records
2. Record relevant versions, inputs, tool calls, and execution metadata
3. Change the system prompt during every investigation
4. Store only the final user-facing answer

Correct Answer: 2

Explanation

Reproducibility requires enough information to understand the environment and execution that produced a particular result. Useful metadata can include model version, prompt or configuration version, relevant tool definitions, inputs, tool calls, timestamps, and execution identifiers. Sensitive information should still be minimized or protected appropriately. Storing only the final answer removes much of the context needed to understand why the failure occurred. Versioned traces and structured logs allow developers to compare executions and determine whether a problem was caused by the model, prompt, tool, data, or application logic.

Question 291

A user requests that an agent delete an account immediately. The account contains important financial records. What additional control may be appropriate?

1. Automatic deletion without verification
2. Removal of all audit logs
3. Confirmation or human review before the destructive action
4. Giving the agent unrestricted database permissions

Correct Answer: 3

Explanation

Deleting an account containing important financial records is a high-impact and potentially irreversible operation. Depending on the application’s requirements, the system may require explicit confirmation, stronger authentication, or human review before executing the deletion. The agent should also validate the target account and verify that the user has permission to perform the operation. Audit records should be retained appropriately so the action can be investigated later. Unrestricted database permissions increase risk, while deleting audit logs removes accountability. Strong safeguards should be proportional to the consequences of the requested action.

Question 292

Why is it useful for a tool to have a focused responsibility rather than many unrelated capabilities?

1. It simplifies tool selection, authorization, testing, and auditing
2. It guarantees that the model cannot hallucinate
3. It removes the need for structured parameters
4. It allows the tool to bypass access controls

Correct Answer: 1

Explanation

Focused tools provide clearer boundaries around what an operation is designed to accomplish. A tool dedicated to retrieving order status is easier to describe, authorize, validate, test, and monitor than a broad tool capable of accessing and modifying an entire database. Focused responsibilities also help the model distinguish between similar operations and reduce the chance of invoking a powerful capability unnecessarily. This design does not eliminate model errors, so validation and backend authorization remain important. However, narrower capabilities reduce the potential impact of incorrect tool selection and make system behavior easier to understand.

Question 293

An agent needs to access a payment service. Which credential-management approach is safest?

1. Put the API key in the system prompt
2. Include the API key in every tool response
3. Store the credential securely and keep it outside the model’s visible context
4. Ask the user to provide the API key in every request

Correct Answer: 3

Explanation

Sensitive credentials should be managed by secure application infrastructure rather than exposed directly to the model. Secret-management systems and scoped service credentials can allow the application to authenticate with an external service without revealing the secret to the agent. Putting API keys into prompts, responses, or user messages increases the chance of accidental exposure through logs, outputs, or malicious content. The tool interface should abstract the credential-handling process from the model whenever possible. Access should also be limited to the specific service and operations required by the workflow.

Question 294

An agent must retrieve current inventory before confirming an order. Why should it retrieve fresh inventory information?

1. Inventory levels can change and stale information may cause incorrect decisions
2. Fresh data always makes responses longer
3. Cached data is always more accurate
4. Inventory never changes after it is first recorded

Correct Answer: 1

Explanation

Inventory is a dynamic value that can change because of purchases, cancellations, returns, or other transactions. If an agent relies on stale information, it could confirm an order for an item that is no longer available. Retrieving current inventory before performing a consequential action helps ensure that the decision is based on relevant information. Caching can still be useful when a defined freshness window is acceptable, but the system should understand how quickly the underlying data changes. Freshness requirements should therefore be aligned with the business operation being performed.

Question 295

What should an agent do when a user provides a request that conflicts with a higher-priority system instruction?

1. Follow the user’s request regardless of priority
2. Follow the higher-priority instruction
3. Combine both instructions even if they conflict
4. Ask the user to rewrite the system instruction

Correct Answer: 2

Explanation

AI applications commonly use an instruction hierarchy in which higher-priority instructions establish system-level behavior and constraints. User requests should be followed when they are compatible with those constraints, but they should not override higher-priority instructions. This is particularly important for security, privacy, tool authorization, and application policies. The agent should not reveal protected information or perform prohibited actions simply because the user requests them. Maintaining the instruction hierarchy provides a consistent control boundary and helps prevent attempts to manipulate the system through conflicting lower-priority instructions.

Question 296

Which approach is most appropriate for a tool that changes a customer’s billing plan?

1. Allow any authenticated user to modify any account
2. Skip validation because billing changes are reversible
3. Require appropriate authorization and validate the target account and plan
4. Give the model administrator-level permissions

Correct Answer: 3

Explanation

Billing-plan changes can affect charges and service access, so the application should enforce appropriate authorization and validate the target account and selected plan before making the change. The agent should not be given administrator-level permissions simply to perform this task. Validation can check that the plan exists, is available, and is compatible with the account. Depending on the workflow, explicit confirmation may also be required before a billing change becomes effective. These controls reduce the risk of unauthorized modifications and help ensure that the requested operation matches the user’s actual intent.

Question 297

An agent’s response contains a customer identifier that does not match the expected format. What should the application do?

1. Accept it because the model generated it
2. Validate the identifier before using it in another operation
3. Replace it with a random identifier
4. Use another customer’s identifier

Correct Answer: 2

Explanation

Identifiers should be validated before they are passed into downstream operations. A malformed customer identifier may indicate a model error, an incorrect tool response, or unexpected input. Using it without validation could cause failed requests or, in poorly designed systems, access to the wrong resource. Schema and business-rule validation can confirm the expected format and, where possible, verify that the identifier belongs to an authorized customer context. Model-generated values should never automatically be treated as trusted simply because they appear plausible. Validation provides an important boundary between model output and application actions.

Question 298

Why should an agent report partial completion instead of claiming the entire task succeeded?

1. Users need an accurate understanding of what actually happened
2. Partial completion is always equivalent to success
3. Reporting failures is unnecessary
4. It allows the agent to hide unsuccessful operations

Correct Answer: 1

Explanation

Accurate status reporting is essential when a multi-step task does not fully complete. If some operations succeeded while another failed, the user should know which actions were completed and which remain unresolved. Claiming full success could cause the user to believe that all requested changes occurred, potentially leading to incorrect follow-up actions. A structured status model can distinguish completed, failed, pending, and skipped steps. Clear partial-completion reporting also helps support recovery, allowing the user or application to retry only the necessary part rather than repeating the entire workflow unnecessarily.

Question 299

An agent receives a request to send an email to a customer, but the recipient address is missing. What should it do?

1. Guess the recipient’s email address
2. Use the first address found in the database
3. Ask for or retrieve the recipient address through an authorized method
4. Send the email without a recipient

Correct Answer: 3

Explanation

A recipient address is a required parameter for sending an email. The agent should obtain the correct address through an authorized method or ask the user to provide it rather than guessing. Choosing an arbitrary address could disclose information to the wrong person, while sending without a valid recipient is technically invalid. Because email creates an external side effect, recipient validation is particularly important. The application should also enforce relevant communication policies and permissions before sending. Clear parameter requirements and structured validation help prevent incorrect or unauthorized messages from being delivered.

Question 300

Which practice provides the strongest foundation for maintaining a reliable agent system as prompts, models, and tools change?

1. Avoid all testing after deployment
2. Use versioning, regression evaluations, structured logging, and controlled changes
3. Allow every component to change independently without records
4. Depend only on manual user feedback

Correct Answer: 2

Explanation

Reliable agent systems require controlled change management because modifications to prompts, models, tools, schemas, or workflows can alter behavior in unexpected ways. Versioning makes changes traceable, while regression evaluations help identify behavioral differences before deployment. Structured logging provides evidence for diagnosing problems after deployment, and controlled releases reduce the impact of unexpected issues. Manual feedback can be valuable but is not sufficient by itself because it may miss edge cases and safety regressions. Combining these practices creates a stronger foundation for monitoring, debugging, evaluating, and continuously improving agent behavior.