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

 

Question 161

An AI agent needs to access a customer’s profile and update the customer’s phone number. Which approach BEST separates the two operations securely?

1. Use one unrestricted tool for both operations.
2. Use separate read and update tools with appropriate permissions.
3. Give the agent administrator access to the customer database.
4. Allow the model to bypass authorization for updates.

Correct Answer: 2

Explanation

Separating read and write operations into different tools makes permissions easier to control and audit. A profile lookup tool can be read-only, while the update tool can require stronger authorization and validation. This follows the principle of least privilege because the agent receives only the capability needed for each specific operation. A single unrestricted tool can make accidental modifications more likely and complicate security controls. Administrator access is unnecessary when the task can be completed through narrowly scoped permissions. Separate tools also make monitoring clearer because developers can distinguish information retrieval from state-changing actions.

Question 162

An agent receives a request to refund a purchase, but the refund amount is not specified. What should the agent do?

1. Refund the entire purchase automatically.
2. Refund the smallest possible amount.
3. Ask the user to clarify the intended refund amount.
4. Choose an amount based on previous refunds.

Correct Answer: 3

Explanation

When a required parameter for a financial action is missing, the agent should request clarification instead of making an unsupported assumption. A refund is a consequential external action, so guessing the amount could create financial errors. The application should also validate that the requested amount is within the allowable refund limit and that the requester is authorized to initiate the refund. If the business workflow has a predefined refund policy, the agent can explain the available options. Explicit clarification preserves user intent and prevents the agent from making a financial decision that the user did not actually authorize.

Question 163

Which mechanism can help ensure that a tool is called only by authorized parts of an application?

1. Access-control checks enforced by the application or service.
2. A longer tool description.
3. A larger context window.
4. A more detailed user prompt.

Correct Answer: 1

Explanation

Authorization should be enforced through application-level or service-level access controls rather than relying solely on prompts or tool descriptions. These controls can determine whether a particular user, service, role, or workflow is permitted to perform an operation. For example, a read-only support agent may be permitted to retrieve order information but not issue refunds. Tool descriptions help the model understand intended use, but they are not security boundaries. Strong authorization checks remain effective even if the model makes an incorrect decision or receives malicious instructions. This provides a more reliable security foundation for agentic applications.

Question 164

An agent retrieves a document containing confidential information that is unrelated to the user’s request. What should the workflow generally do?

1. Include all confidential information in the final response.
2. Pass the confidential information to every downstream tool.
3. Minimize or exclude unrelated sensitive information from further processing.
4. Store the entire document permanently.

Correct Answer: 3

Explanation

Information that is unrelated to the user’s task should generally not be unnecessarily exposed or propagated through the workflow. Data minimization reduces the number of components that can access sensitive information and limits the impact of potential leaks. The application can filter relevant sections, redact sensitive fields, or restrict which content is passed to downstream tools. Storing confidential information permanently without a legitimate need can also increase privacy and security risks. The goal is to provide the model with sufficient information to complete the task while avoiding unnecessary exposure of unrelated sensitive data.

Question 165

What is the MAIN reason an agent workflow should distinguish between a completed task and a partially completed task?

1. To ensure the user receives an accurate status of what actually happened.
2. To make responses longer.
3. To avoid recording tool results.
4. To allow the agent to claim success earlier.

Correct Answer: 1

Explanation

A multi-step workflow can succeed in some operations while failing in others. If the agent reports the entire task as successful when only part of it completed, the user may make decisions based on incorrect information. Explicit workflow states such as pending, partially completed, successful, and failed allow the application to accurately communicate progress. For example, an order may be updated successfully while a notification email fails. Reporting these outcomes separately is more useful than claiming total success. Accurate completion states also improve auditing, debugging, and recovery from failed steps.

Question 166

An agent uses a third-party API to retrieve customer information. Which practice BEST reduces unnecessary exposure of customer data?

1. Send the customer’s entire account history with every API request.
2. Send only the minimum information required by the API.
3. Include unrelated customer records for additional context.
4. Disable all API validation.

Correct Answer: 2

Explanation

When information is sent to a third-party service, the application should minimize the data shared with that service. If the API only requires a customer identifier, there is generally no reason to send the customer’s full account history or unrelated personal information. Data minimization reduces privacy exposure and limits the consequences of a potential security incident. The application should also consider contractual, regulatory, and organizational requirements governing third-party data sharing. Carefully scoped API requests provide the required functionality without unnecessarily expanding the amount of sensitive information exposed outside the core system.

Question 167

An agent repeatedly generates invalid tool parameters despite having a valid schema. What should developers investigate FIRST?

1. Whether the tool description and parameter requirements are clear and unambiguous.
2. Whether all authorization controls should be removed.
3. Whether the database should be deleted.
4. Whether the tool should accept arbitrary parameters.

Correct Answer: 1

Explanation

Repeated invalid tool calls may indicate that the model does not clearly understand the tool’s purpose or parameter requirements. Developers should review the tool description, schema, required fields, allowed values, examples, and error messages. The interface should make important constraints explicit and consistent. Application-side validation should remain in place even after improving the description. Removing authorization or accepting arbitrary parameters would increase risk rather than solve the underlying problem. Developers can also use evaluation traces to identify recurring failure patterns and determine whether the issue comes from tool selection, parameter generation, or workflow design.

Question 168

Which situation is MOST appropriate for requiring human review in an agent workflow?

1. Formatting a user’s text.
2. Retrieving a public document.
3. Performing a high-impact irreversible operation with uncertain consequences.
4. Sorting a list alphabetically.

Correct Answer: 3

Explanation

Human review can be valuable when an action has significant consequences, is difficult to reverse, or contains unresolved uncertainty. Examples may include irreversible deletion, large financial transactions, sensitive account changes, or decisions involving substantial external impact. Lower-risk tasks such as formatting text or retrieving public information usually require fewer safeguards. Human review should be incorporated according to the risk of the operation rather than applied indiscriminately. The objective is to provide appropriate oversight where automated execution could cause significant harm while allowing routine, low-risk tasks to remain efficient.

Question 169

An agent receives an error indicating that a requested resource does not exist. What should it generally do?

1. Repeatedly retry the exact same request indefinitely.
2. Treat the missing resource as successfully retrieved.
3. Handle the not-found condition explicitly and request clarification or take an approved alternative.
4. Create a resource without authorization.

Correct Answer: 3

Explanation

A not-found error generally indicates that the requested resource could not be located, so the workflow should handle that state explicitly. The agent may ask the user to verify the identifier, search using an approved alternative, or explain that the resource does not exist. Repeating the same request indefinitely is unlikely to solve a permanent not-found condition. Creating a new resource without authorization could also produce unintended side effects. Clear error categories allow the agent to distinguish permanent conditions such as missing resources from temporary failures that may be appropriate for retry.

Question 170

What is the PRIMARY benefit of using focused tool descriptions rather than one generic description for all tools?

1. It helps the agent distinguish which tool is appropriate for a specific task.
2. It guarantees perfect tool selection.
3. It eliminates the need for schemas.
4. It allows every tool to perform the same operation.

Correct Answer: 1

Explanation

Focused tool descriptions make it easier for an agent to understand the purpose and appropriate use of each available capability. If several tools are available, clear descriptions help distinguish their responsibilities and reduce ambiguous selection. For example, separate descriptions for searching orders, updating orders, and issuing refunds are easier to reason about than a single generic tool description covering all operations. Tool descriptions do not guarantee perfect behavior, so application-level validation and authorization remain necessary. Clear interfaces combined with focused descriptions generally improve tool selection and make workflows easier to maintain.

Question 171

An agent is asked to modify a user’s email address, but the request comes from an unauthenticated session. What should happen?

1. Perform the update because the user supplied the new address.
2. Require appropriate authentication before modifying the account.
3. Change the email address temporarily.
4. Send the new address to another customer for verification.

Correct Answer: 2

Explanation

Account modifications should require appropriate authentication and authorization. An unauthenticated request does not establish that the requester controls the account. The application should therefore require the approved authentication flow before allowing the agent to change account information. Depending on the system, additional verification may be required for sensitive account changes. Simply accepting the new email address does not establish ownership. Sending account information to another customer would create an obvious privacy problem. Authentication and authorization should be enforced by the application rather than relying on the agent to determine whether a request appears legitimate.

Question 172

Which practice BEST improves reproducibility when debugging an agent failure?

1. Record relevant workflow versions, tool calls, inputs, outputs, and timestamps.
2. Delete logs after every execution.
3. Change the prompt during every test.
4. Record only the final answer.

Correct Answer: 1

Explanation

Reproducibility requires enough information to understand the conditions under which an agent produced a particular result. Useful information can include the workflow or prompt version, tool definitions, relevant inputs, tool calls, outputs, timestamps, and error states. Sensitive data should be minimized or redacted where appropriate. Recording only the final answer makes it difficult to determine whether a problem originated in retrieval, tool selection, parameter generation, external services, or the final response. Versioned traces and structured logs therefore provide a stronger foundation for debugging and regression analysis.

Question 173

An agent has to process information from an untrusted website and then call an internal administrative tool. What is the MOST important security consideration?

1. Allow the website content to control the administrative tool.
2. Treat website instructions as untrusted and enforce authorization before any internal action.
3. Give the website temporary administrator permissions.
4. Disable validation because the website is public.

Correct Answer: 2

Explanation

Untrusted external content should not automatically gain authority over internal tools. Websites can contain instructions designed to manipulate an agent into taking actions that were never authorized by the application or user. The workflow should separate retrieved data from trusted instructions and enforce authorization before any administrative operation. Even if the model interprets the webpage as instructing it to perform an action, the application should validate whether that action is actually permitted. This defense helps protect internal systems from prompt injection and prevents external content from bypassing established security boundaries.

Question 174

Why should an agent’s writable tool permissions generally be narrower than its read permissions when possible?

1. Write operations can create persistent side effects and therefore carry greater risk.
2. Read operations are always more dangerous than writes.
3. Narrow permissions prevent all errors.
4. Write permissions are unnecessary in every application.

Correct Answer: 1

Explanation

Write operations can change persistent data, trigger external actions, or create financial and operational consequences. Because these effects may be difficult to reverse, write permissions should generally be narrowly scoped. An agent might need broad access to retrieve information but only limited access to modify records. This supports least privilege and reduces the impact of incorrect tool selection, prompt injection, or model errors. Narrow permissions do not eliminate all risks, so validation, authorization, confirmation, and auditing may still be required. The appropriate permission scope depends on the actual task and potential impact.

Question 175

An agent needs to call a tool that requires a date and time. The user says, “Schedule it for tomorrow afternoon.” What should the agent do if the exact time is required?

1. Choose an arbitrary time without telling the user.
2. Ask for the specific time or use an explicitly defined application default.
3. Schedule the event at midnight.
4. Schedule multiple events for different times.

Correct Answer: 2

Explanation

If the tool requires an exact time and the user’s request only specifies a broad period, the agent should either ask for clarification or use a clearly defined application default if such a default is authorized. Choosing an arbitrary time without communicating it can lead to unwanted commitments. Scheduling multiple events to cover the uncertainty creates unnecessary side effects. The workflow should also correctly interpret relative dates such as “tomorrow” using the relevant timezone and current date. Explicit handling of ambiguous temporal information improves reliability and prevents accidental scheduling.

Question 176

Which practice helps an agent application detect whether a new model version causes previously successful tasks to fail?

1. Running a versioned regression test suite against the new model.
2. Removing old test cases.
3. Evaluating only newly added features.
4. Measuring only model response length.

Correct Answer: 1

Explanation

A regression suite contains previously defined tasks that represent important expected behaviors. Running the same suite against a new model version allows developers to compare results and identify changes in performance or behavior. Tests can cover tool use, retrieval, structured output, authorization boundaries, error handling, and other application-specific requirements. Removing old tests eliminates the baseline needed to detect regressions. Response length is not a reliable indicator of correctness or safety. Versioned evaluations therefore provide a repeatable way to monitor changes as models and agent workflows evolve.

Question 177

An agent is instructed to provide a structured JSON response. The generated output contains invalid JSON. What should the application do?

1. Parse it as valid JSON anyway.
2. Validate the output and handle the formatting failure appropriately.
3. Treat the invalid output as a successful database update.
4. Remove the output schema permanently.

Correct Answer: 2

Explanation

Structured outputs should be validated before the application relies on them. If the model generates invalid JSON, the application can reject the result, request a corrected response, or follow another approved recovery path. Depending on the API or implementation, structured-output mechanisms can further constrain the expected format. Treating invalid data as valid can cause downstream failures or incorrect actions. Removing the schema would make the interface less predictable rather than solving the immediate problem. Validation provides a reliable boundary between model-generated content and application logic.

Question 178

An agent successfully retrieves a customer’s order but fails when attempting to send the requested notification email. What is the MOST accurate final status?

1. Fully completed.
2. Failed before any work occurred.
3. Partially completed, with the email step unsuccessful.
4. Completed successfully because the order was retrieved.

Correct Answer: 3

Explanation

The workflow should distinguish between successful and unsuccessful steps. In this case, the order retrieval succeeded, but the notification email failed. Reporting the task as fully completed would incorrectly imply that the customer was notified. The application should communicate the partial completion and, where appropriate, provide an option to retry the failed step. Explicit workflow states make it easier to recover from failures and keep the user informed. This also improves auditability because developers can determine which operations succeeded and which require additional action.

Question 179

What is the BEST reason to use an audit trail for high-impact agent actions?

1. It provides a record of important actions for accountability and investigation.
2. It guarantees that the agent cannot make mistakes.
3. It replaces authentication.
4. It allows unrestricted access to sensitive data.

Correct Answer: 1

Explanation

An audit trail records important actions performed by an agent or application, helping organizations understand what happened and when. Depending on the system, records may include the actor or service identity, operation, target, timestamp, result, and relevant identifiers. Audit trails support accountability, troubleshooting, security investigations, and compliance requirements where applicable. They do not prevent mistakes by themselves and cannot replace authorization or authentication. Logging should also follow data-minimization principles so that unnecessary sensitive information is not retained. A well-designed audit trail provides useful evidence without becoming an additional privacy risk.

Question 180

An agent has several available tools, but only one tool can safely perform the requested operation. What should guide the tool-selection decision?

1. The tool with the longest description.
2. The tool that requires the most permissions.
3. The tool whose defined purpose, parameters, and authorization match the requested operation.
4. A random tool selected from the available options.

Correct Answer: 3

Explanation

Tool selection should be based on the tool’s defined purpose, required parameters, and applicable authorization rather than arbitrary characteristics. A focused tool that directly matches the requested operation is generally easier to use safely than a broader tool with unnecessary capabilities. The application should still validate the generated parameters and enforce authorization before execution. Selecting a tool simply because it has a long description or requires more permissions does not improve correctness. Good tool architecture makes responsibilities clear, allowing the agent to identify the appropriate capability while the application maintains final control over what can actually be executed.