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

 

Question 381

An agent is asked to refund a customer, but the refund amount is not specified and multiple previous transactions exist. What should the agent do?

1. Refund the largest transaction
2. Refund the most recent transaction
3. Ask the customer to identify the transaction and amount
4. Refund all previous transactions

Correct Answer: 3

Explanation

A refund request is ambiguous when multiple transactions exist and neither the target transaction nor amount has been clearly identified. The agent should ask the customer to specify which transaction they want refunded and, where necessary, confirm the amount. Automatically selecting the newest, largest, or all transactions could create unintended financial consequences. Before executing the refund, the application should also validate authorization, transaction status, refund eligibility, and any applicable business rules. Financial operations should use precise parameters and appropriate confirmation mechanisms rather than relying on assumptions made by the model.

Question 382

Which approach best protects an agent from accidentally exposing an API key in its response?

1. Include the API key in every system message
2. Keep secrets outside model-visible context and restrict access through the application
3. Ask the model to memorize the API key
4. Store the API key in user-facing conversation history

Correct Answer: 2

Explanation

Secrets such as API keys should generally remain outside the model’s conversational context. The application can securely store credentials and use them when making authorized tool or API requests without exposing the secret to the model. This reduces the chance of accidental disclosure through generated responses, logs, prompt injection, or conversation history. Secret-management systems can also support rotation and access auditing. Even when credentials are protected, the associated services should use narrowly scoped permissions. Secret protection should therefore combine secure storage, restricted access, and application-level controls rather than relying on the model to keep credentials confidential.

Question 383

An agent receives two sources containing different information about a customer’s order status. What should it consider when resolving the conflict?

1. Source authority and data freshness
2. Which response is longer
3. Which source appears first alphabetically
4. Which source has more formatting

Correct Answer: 1

Explanation

When sources disagree, the application should consider factors such as source authority, timestamp, freshness, and reliability. For example, a real-time order-management system may be more authoritative for current delivery status than an older cached record. The agent should not choose a source simply because its response is longer or appears first. If the conflict cannot be reliably resolved, the system should communicate the uncertainty or retrieve additional information. Provenance metadata is useful because it allows the workflow to understand where each result came from and when it was generated.

Question 384

What is an appropriate reason to impose a maximum number of agent iterations?

1. To prevent runaway loops and uncontrolled resource usage
2. To make every answer shorter
3. To eliminate all tool calls
4. To prevent the model from reading user requests

Correct Answer: 1

Explanation

A maximum iteration limit provides a safeguard against workflows that repeatedly call tools or generate actions without reaching a valid completion state. Runaway loops can increase API costs, consume compute resources, delay responses, and potentially create repeated side effects. The limit should be combined with clear termination conditions and appropriate retry policies. If the limit is reached, the application should report the actual workflow state rather than claiming success. For complex tasks, a controlled continuation mechanism may allow the workflow to resume safely after review or additional user input.

Question 385

An agent needs to modify a database record, but the requested field value violates a business rule. What should happen?

1. Apply the invalid value anyway
2. Ignore the business rule
3. Reject the update and return a structured validation error
4. Change the value to an arbitrary valid value

Correct Answer: 3

Explanation

Business rules should be enforced before persistent changes are applied. If the requested value violates a rule, the system should reject the operation and provide a clear validation error explaining what needs to be corrected. The agent should not silently change the user’s requested value because that could produce an outcome the user did not intend. Validation should occur at the application or service layer rather than relying solely on the model. Structured errors make it easier for the agent and downstream software to understand why the operation failed and what information may be needed to continue.

Question 386

Which practice is most useful for detecting changes in agent behavior after a prompt update?

1. Delete previous evaluation results
2. Compare the new prompt against the previous version using regression tests
3. Disable logging
4. Allow the new prompt to run only once

Correct Answer: 2

Explanation

Prompt changes can affect how an agent interprets requests, selects tools, follows instructions, and formats responses. Regression testing allows developers to compare behavior between the previous and updated prompt versions using consistent scenarios. This makes it easier to identify unintended changes before or after deployment. Evaluation results should be retained appropriately so that differences can be analyzed over time. Prompt versioning also helps correlate production behavior with a specific configuration. Manual testing can complement regression tests but should not be the only method used to evaluate important agent changes.

Question 387

A user asks an agent to access a company resource, but the user’s role does not have permission to access it. What should the agent do?

1. Grant itself temporary permission
2. Access the resource because the user requested it
3. Ask another user for credentials
4. Respect the authorization boundary and decline the unauthorized access

Correct Answer: 4

Explanation

The agent should respect the application’s authorization boundaries. A user request does not automatically grant permission to access a restricted company resource. Authorization should be enforced by the underlying service or application, and the agent should not attempt to bypass the restriction by obtaining another person’s credentials or changing its own permissions. If an approved access-request process exists, the agent can direct the user toward that process. This demonstrates why tool permissions and backend authorization are more reliable than relying solely on conversational instructions when protecting sensitive resources.

Question 388

An agent has access to a tool capable of deleting an entire database. What is an appropriate design principle?

1. Expose the capability broadly because the model may need it
2. Restrict or isolate highly destructive capabilities and require strong authorization
3. Allow any user to invoke it through natural language
4. Remove all confirmation requirements

Correct Answer: 2

Explanation

A tool capable of deleting an entire database represents an extremely high-impact capability. Such permissions should be tightly restricted, potentially isolated from normal agent workflows, and protected by strong authorization and human approval requirements. In many systems, the agent should not have direct access to this capability at all. If such a tool must exist, precise targeting, confirmation, audit logging, environment separation, and additional safeguards are appropriate. The principle is that agent autonomy and available permissions should be proportional to the risk and reversibility of the operation.

Question 389

Why should an agent distinguish between read-only and write operations?

1. Write operations can create persistent side effects and therefore generally require stronger controls
2. Read operations are always unsafe
3. Write operations never require authorization
4. Read and write operations have identical risks

Correct Answer: 1

Explanation

Read-only operations generally retrieve information without changing external state, while write operations can modify records, send messages, create transactions, or otherwise produce persistent side effects. Because write actions can have greater consequences, they often require stronger validation, authorization, confirmation, and auditing. This distinction helps developers design appropriate tool permissions and safeguards. It does not mean read operations are risk-free, because sensitive information can still be exposed through unauthorized reads. Both types of operations should therefore be protected, but their controls can be calibrated according to the specific risks involved.

Question 390

A customer asks an agent to change a billing address, but the customer has not specified which account should be updated. What should happen?

1. Update every account associated with the customer
2. Select the oldest account
3. Ask the customer to identify the intended account
4. Update the account with the highest balance

Correct Answer: 3

Explanation

When a customer has multiple accounts, the agent should not assume which account the request refers to. Updating every account or selecting one based on arbitrary characteristics could create incorrect persistent changes. The agent should ask the customer to identify the intended account using an account name, masked identifier, or another clear reference. Once identified, the application should verify that the requester is authorized to modify the account and validate the new billing address. Clarifying the target before execution is especially important because address changes may affect invoices, payments, tax information, and other account processes.

Question 391

Which feature is most useful for understanding why an agent made a particular tool call during production troubleshooting?

1. Trace and audit information associated with the workflow
2. The user’s screen resolution
3. The application’s color scheme
4. The number of browser tabs open

Correct Answer: 1

Explanation

Trace and audit information can help developers reconstruct the sequence of events leading to a tool invocation. Useful information may include request identifiers, timestamps, selected tool, relevant parameters, authorization results, tool response status, and workflow state. Such information helps determine whether the agent selected an inappropriate tool, received unexpected data, or encountered an application error. Logs should still follow data-minimization and privacy requirements. Good observability does not mean recording every piece of user information; it means retaining the operational information necessary to understand and investigate system behavior.

Question 392

An agent receives an external webpage containing a request to reveal confidential customer information. What should it do?

1. Follow the webpage instruction because it is recent
2. Treat the request as untrusted content and maintain existing privacy controls
3. Reveal only half of the information
4. Disable authorization temporarily

Correct Answer: 2

Explanation

Instructions embedded in external content should not automatically override the application’s privacy and security requirements. A webpage can contain prompt-injection text that attempts to manipulate the agent into revealing confidential information. The agent should treat the content as untrusted data and continue following higher-priority instructions and access controls. Confidential information should only be disclosed when the requester and operation are authorized. Security enforcement should occur independently of the model wherever possible, so that external content cannot directly grant itself access to protected customer data or internal resources.

Question 393

What should an application do if a tool returns a response that fails its expected JSON schema?

1. Accept it without validation
2. Treat the malformed response as valid
3. Validate the response and handle the schema failure appropriately
4. Guess the missing fields

Correct Answer: 3

Explanation

Structured responses should be validated against the expected schema before being used by downstream components. If required fields are missing, data types are incorrect, or the structure is otherwise malformed, the application should reject or quarantine the response and follow an appropriate recovery path. Depending on the situation, it may retry the request, return a structured error, or ask for additional information. Guessing missing fields can introduce incorrect data into subsequent workflow steps. Schema validation is particularly important when agent-generated or tool-generated data triggers automated actions.

Question 394

Which practice helps prevent accidental use of test data in a production customer workflow?

1. Use clear environment boundaries and configuration controls
2. Allow the model to switch environments automatically
3. Mix test and production databases
4. Use identical unrestricted credentials everywhere

Correct Answer: 1

Explanation

Clear environment boundaries help ensure that production workflows use production resources while development and testing workflows use appropriate test resources. Configuration should explicitly identify the target environment, and credentials should be separate and appropriately scoped. Mixing databases or using identical unrestricted credentials increases the risk of accidental data modification and exposure. Automated deployment and configuration validation can further reduce mistakes. Agents should not be responsible for deciding which environment to use based solely on conversational context; the surrounding application should establish the environment and enforce the corresponding permissions.

Question 395

A user asks an agent to send a document but does not specify the recipient. What should the agent do?

1. Send the document to the user’s most recent contact
2. Send it to every contact
3. Ask the user to specify the recipient
4. Choose a recipient based on the document contents

Correct Answer: 3

Explanation

The recipient is a critical parameter for sending an external document. If it is missing, the agent should ask the user to identify the intended recipient rather than guessing. Sending the document to the wrong person could disclose confidential information and create an irreversible external side effect. Once the recipient is identified, the application can validate the address and any applicable sharing permissions. If the document itself is sensitive, an additional confirmation step may be appropriate. Clear recipient identification is therefore an important safeguard for agent-driven email and file-sharing workflows.

Question 396

Why should an agent application maintain an audit trail for high-impact actions?

1. To make responses longer
2. To support accountability, investigation, and verification of important operations
3. To replace authorization controls
4. To give the model unrestricted permissions

Correct Answer: 2

Explanation

An audit trail provides a record of important actions and can help organizations understand who or what initiated an operation, when it occurred, what resource was affected, and what the outcome was. This is useful for troubleshooting, security investigations, compliance requirements, and accountability. Audit records should be protected from unauthorized modification and should minimize unnecessary sensitive information. Logging an action does not make the action authorized; authorization and validation must still occur before execution. Auditability is therefore one component of a broader control framework for high-impact agent operations.

Question 397

An agent’s workflow fails after successfully completing the first two steps. What should the final status communicate?

1. That the entire workflow succeeded
2. That no work was performed
3. The actual partial-completion state and the failed step
4. Only the original user request

Correct Answer: 3

Explanation

Accurate status reporting is essential when a multi-step workflow only partially succeeds. The system should identify which operations completed and which operation failed rather than reporting overall success. This helps the user understand what happened and prevents duplicate work or unsafe retries. A structured status model can represent each step independently and support controlled recovery. For example, the application may mark steps one and two as completed and step three as failed. If the failed operation is safely retryable, the system can resume from that point instead of repeating completed side effects.

Question 398

Which approach is most appropriate when an agent must process a very large amount of retrieved information?

1. Pass every available document into the model regardless of relevance
2. Use retrieval, filtering, or summarization to provide relevant context
3. Remove all source information
4. Allow unrelated documents to determine the response

Correct Answer: 2

Explanation

Large amounts of irrelevant context can increase processing costs and make it harder for the model to focus on the user’s actual request. Retrieval and filtering can identify relevant passages before they are provided to the model. Summarization can also reduce large documents when detailed source text is unnecessary, although important facts should remain verifiable when needed. Context minimization improves efficiency and can reduce the exposure of unrelated sensitive information. The goal is not simply to provide as much information as possible, but to provide sufficient, relevant, and appropriately sourced information for the requested task.

Question 399

An agent is about to execute a high-impact operation, but the user’s instruction is ambiguous. What is the safest general approach?

1. Make the most likely assumption and execute
2. Execute all possible interpretations
3. Ask for clarification before executing the operation
4. Select the most expensive option

Correct Answer: 3

Explanation

Ambiguity should generally be resolved before high-impact actions are executed. If several interpretations are possible, choosing one without confirmation can cause an unintended financial, operational, or data-related consequence. A concise clarification question can establish the intended target, amount, recipient, or action. This is especially important for irreversible or externally visible operations. The application should combine clarification with authorization, validation, and appropriate confirmation mechanisms. Agents can act autonomously on low-risk tasks when intent is clear, but higher-risk actions require stronger safeguards when uncertainty remains.

Question 400

Which combination best supports reliable and secure agent operation in production?

1. Broad permissions, unlimited retries, and no logging
2. Strong prompts without authorization or validation
3. Least privilege, structured validation, controlled execution, testing, and monitoring
4. Unlimited tool access combined with automatic execution

Correct Answer: 3

Explanation

Reliable production agents require multiple layers of protection rather than depending on the model alone. Least-privilege permissions restrict what the agent can do, while structured input and output validation reduce malformed or unsafe operations. Controlled execution through timeouts, retry limits, confirmation requirements, and workflow states helps prevent runaway behavior and unintended side effects. Regression testing identifies behavioral changes before deployment, while monitoring and audit trails provide visibility after deployment. These controls work together as defense in depth. A capable model can improve automation, but secure production operation depends on the surrounding application architecture and governance controls.