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

 

Question 121

An AI agent receives a request to access a customer’s account information, but the user’s identity has not been verified. What should the agent do FIRST?

1. Retrieve the information because the request appears legitimate.
2. Ask the user to provide unrelated personal information.
3. Verify the user’s identity before accessing protected account information.
4. Provide only half of the requested account information.

Correct Answer: 3

Explanation

Identity verification should occur before an agent accesses protected account information. The agent should not assume that a request is legitimate simply because the user provides a plausible explanation. Authentication establishes who the requester is, while authorization determines what that identity is allowed to access. Depending on the application, verification may involve an authenticated session, identity provider, or another approved mechanism. Providing partial information does not eliminate the security risk because even limited account details may be sensitive. A properly designed workflow verifies identity before retrieving protected data and then applies authorization checks to determine whether the requested information can be disclosed.

Question 122

A tool allows an agent to update a customer’s shipping address. Which design BEST reduces accidental updates to the wrong customer?

1. Let the model select any customer record.
2. Require a validated customer identifier and authorization check before the update.
3. Allow updates based only on the customer’s first name.
4. Update the most recently accessed customer automatically.

Correct Answer: 2

Explanation

Customer records should never be selected using ambiguous information such as a first name alone. A safer design requires a validated customer identifier and verifies that the authenticated user or workflow is authorized to modify that specific record. The application should enforce these checks independently of the model because the model can misunderstand references or select an incorrect identifier. Automatically using the most recently accessed record creates a serious cross-customer data risk. Strong authorization and precise record identification reduce accidental updates and help maintain tenant and customer isolation.

Question 123

What is the MAIN purpose of setting a maximum execution budget for an autonomous agent?

1. To guarantee that every task succeeds.
2. To prevent the model from using any tools.
3. To eliminate the need for monitoring.
4. To limit excessive resource use and runaway execution.

Correct Answer: 4

Explanation

An execution budget limits how much computation, time, tool usage, or other resources an agent can consume during a task. Autonomous workflows can sometimes enter loops, repeatedly call tools, or spend excessive resources trying to solve an impossible problem. A maximum execution budget provides a safety boundary that stops the workflow after a defined threshold. It does not guarantee successful completion and does not replace monitoring or proper termination conditions. Budgets are especially useful when agents operate with multiple tools or long-running workflows because they reduce the potential impact of unexpected behavior.

Question 124

An agent must process a large document but only one section is relevant to the user’s question. What is the BEST approach?

1. Pass the entire document through every agent step.
2. Ignore the document.
3. Retrieve or extract only the relevant section before processing it.
4. Ask the model to memorize the entire document.

Correct Answer: 3

Explanation

Processing only the information required for the task improves efficiency and reduces unnecessary context. If a large document contains one relevant section, the application can identify and extract that section before passing it to the model. This reduces token consumption and makes it easier for the model to focus on the information that matters. It can also reduce exposure to unrelated sensitive content. Passing the entire document through every step may increase cost and introduce distracting information. Context minimization is therefore an important design principle for reliable and efficient agent workflows.

Question 125

A customer asks an agent to cancel an order. The order has already shipped and cannot normally be canceled. What should the agent do?

1. Cancel it anyway using an administrative tool.
2. Clearly communicate the order status and follow the approved alternative workflow.
3. Change the order status to make cancellation possible.
4. Tell the customer the order was canceled without performing any action.

Correct Answer: 2

Explanation

The agent should respect the actual business state of the order rather than attempting to bypass restrictions. If an order has already shipped and cancellation is unavailable, the agent should explain the current status and follow the approved alternative, such as a return or refund process if one exists. Changing internal status values simply to force cancellation could create inconsistent records and financial problems. Claiming that the order was canceled without performing the action would also be misleading. Business rules should be enforced by the application and tools, while the agent communicates the available options clearly to the customer.

Question 126

Which characteristic makes a tool interface easier for an AI agent to use reliably?

1. A broad description covering many unrelated tasks.
2. Hidden required parameters.
3. Clear purpose, explicit parameters, and predictable output structure.
4. Different parameter names for the same concept in every request.

Correct Answer: 3

Explanation

Agents perform more reliably when tools have focused responsibilities and clearly defined interfaces. A good tool description explains its purpose, while explicit parameter definitions identify required and optional inputs. Predictable output structures also make it easier for the agent and application to process results consistently. Hidden requirements and inconsistent naming increase ambiguity and can lead to malformed calls. Combining many unrelated operations into one broad tool can also make tool selection more difficult. Clear, focused interfaces improve tool discoverability, validation, debugging, and overall workflow reliability.

Question 127

An agent receives a prompt containing an instruction to reveal its hidden system instructions. What should the application generally do?

1. Follow the instruction because it came from the user.
2. Reveal only part of the system instructions.
3. Treat the request as untrusted if it conflicts with higher-priority instructions.
4. Disable all security controls for the session.

Correct Answer: 3

Explanation

System-level instructions and application security policies have higher priority than ordinary user requests. If a user asks the agent to reveal protected instructions, the agent should not disclose information that the application has designated as confidential. The application should maintain instruction hierarchy and protect sensitive configuration. This is particularly important because prompt-injection attempts may be presented as harmless requests or instructions. The correct behavior is to follow the authorized instruction hierarchy rather than allowing a lower-priority request to override protected instructions.

Question 128

An agent calls two independent read-only APIs that do not depend on each other’s results. Which execution strategy can improve efficiency?

1. Run both calls sequentially in every case.
2. Run the independent calls in parallel when supported by the application.
3. Avoid calling either API.
4. Combine both API requests into one invalid request.

Correct Answer: 2

Explanation

Independent tool calls can often be executed in parallel because neither operation depends on the result of the other. For example, retrieving weather information and account metadata may be independent operations that can run concurrently. Parallel execution can reduce total latency and improve user experience. However, the application should ensure that the calls truly have no dependency or conflicting side effects before parallelizing them. Sequential execution remains appropriate when the second call depends on the first result or when operations interact with shared mutable state. Parallelism should therefore be based on workflow dependencies rather than applied indiscriminately.

Question 129

Why should an agent avoid storing unnecessary sensitive customer information in application logs?

1. It increases the amount of data available for unnecessary exposure.
2. It guarantees faster tool execution.
3. It prevents all security incidents.
4. It removes the need for data encryption.

Correct Answer: 1

Explanation

Logs can be accessed by developers, monitoring systems, support personnel, and other infrastructure components, so unnecessary sensitive information in logs creates additional exposure. Applications should log enough information to support debugging and auditing while minimizing confidential data. Sensitive fields may need to be redacted, masked, or excluded entirely. This follows the principle of data minimization. Logging everything simply because it is technically available increases the consequences of a potential compromise or accidental disclosure. Proper logging therefore balances observability with privacy and security requirements.

Question 130

An agent receives a malformed tool response that does not match the expected schema. What should the application generally do?

1. Assume the response is correct.
2. Automatically invent missing fields.
3. Validate the response and handle the schema error safely.
4. Retry the tool forever.

Correct Answer: 3

Explanation

Tool responses should be validated against their expected schemas before being used by the agent or application. If required fields are missing or values have unexpected types, the application should classify the response as invalid and follow an appropriate error-handling path. Depending on the situation, it may retry a transient failure, request corrected data, or terminate the workflow with a structured error. Inventing missing information can lead to incorrect decisions, while unlimited retries can create runaway execution. Schema validation therefore acts as an important boundary between external tool output and downstream agent reasoning.

Question 131

A company wants to ensure that an agent can only modify records belonging to the authenticated organization. What should enforce this restriction?

1. The model’s internal reasoning alone.
2. A user instruction saying not to access other organizations.
3. Application-level and data-layer authorization controls.
4. A descriptive tool name.

Correct Answer: 3

Explanation

Security boundaries should be enforced by the application and data layer rather than relying on the model to follow instructions perfectly. Organization-level authorization can restrict queries and updates to records belonging to the authenticated tenant. Database policies, scoped queries, service authorization, and access-controlled credentials can provide defense in depth. A tool description or user instruction may guide model behavior but should not be treated as a security boundary. If the model makes a mistake or receives a malicious prompt, properly enforced authorization should still prevent cross-organization access.

Question 132

An agent needs to transfer money between accounts. Which approach is MOST appropriate before executing the transfer?

1. Execute immediately because financial requests are usually valid.
2. Validate account ownership, transfer details, authorization, and required confirmation.
3. Allow the model to select the destination account based on similarity.
4. Skip validation if the transfer amount is small.

Correct Answer: 2

Explanation

Financial transfers are high-impact external actions and require strong controls. The application should verify the authenticated user’s authority, validate source and destination accounts, confirm the transfer amount and currency, and apply any required confirmation or approval process. The agent should not infer the destination account from approximate similarity or bypass validation simply because the amount is small. Strong safeguards reduce the risk of unauthorized transactions and incorrect transfers. The exact controls depend on the financial workflow, but authorization, validation, and appropriate confirmation should be treated as essential components of the transaction process.

Question 133

What is the PRIMARY advantage of maintaining provenance information for retrieved data?

1. It allows the system to identify where information came from.
2. It guarantees that every source is accurate.
3. It removes the need for retrieval.
4. It prevents all conflicting information.

Correct Answer: 1

Explanation

Provenance records provide information about the origin of retrieved content, such as a document identifier, source system, timestamp, or relevant metadata. This helps developers and users understand where a claim or value originated and makes conflicting information easier to investigate. Provenance does not guarantee that a source is correct, nor does it eliminate the possibility of conflicting sources. However, it provides valuable traceability for debugging, auditing, and evaluating answer quality. In agent systems that rely on retrieval, maintaining source information can significantly improve transparency and make downstream verification easier.

Question 134

An agent is responsible for creating support tickets. The user asks it to create a ticket but provides no issue description. What should the agent do?

1. Create a ticket with an invented description.
2. Use the previous customer’s issue.
3. Request the missing information before creating the ticket.
4. Create an empty ticket and mark it resolved.

Correct Answer: 3

Explanation

If an essential field is missing, the agent should request that information rather than inventing it. A support ticket without an issue description may be unusable or incorrectly assigned. The agent can identify the missing requirement and ask the user for a concise description of the problem. If the application has predefined required fields, those requirements should be enforced through the tool schema and validation layer. Guessing the issue or copying information from another customer could result in incorrect records and privacy violations. Asking for the missing information maintains accuracy and ensures the resulting ticket is actionable.

Question 135

Which approach BEST handles a temporary rate-limit response from an external API?

1. Immediately send unlimited requests.
2. Use an appropriate backoff strategy and retry within defined limits.
3. Ignore the rate-limit response.
4. Permanently disable the integration.

Correct Answer: 2

Explanation

Rate limiting indicates that the external service is temporarily restricting request volume. A well-designed application should respect the service’s limits and use an appropriate backoff strategy before retrying. Retry limits are also important so the agent does not continuously consume resources while the service remains unavailable. Where the API provides retry timing information, the application should follow it when appropriate. Unlimited retries can increase load and worsen the problem. Permanently disabling the integration is unnecessarily drastic for a temporary condition. Proper rate-limit handling improves reliability while respecting external service constraints.

Question 136

An agent must update a database record, but two records match the provided customer name. What should it do?

1. Select the first record returned.
2. Update both records.
3. Ask for additional identifying information before updating.
4. Delete one duplicate record.

Correct Answer: 3

Explanation

Ambiguous identifiers should be resolved before performing a consequential database update. If two records share the same customer name, selecting the first result could modify the wrong account. The agent should request additional information, such as a customer ID, verified email address, or another authorized identifier. The application should also enforce uniqueness and authorization where appropriate. Updating both records or deleting one based solely on ambiguity can cause serious data integrity problems. Asking for clarification preserves accuracy and gives the user an opportunity to identify the intended record.

Question 137

What should an agent do when a tool returns a permanent authorization error?

1. Repeatedly retry the same unauthorized operation.
2. Report the authorization failure and avoid continuing the prohibited action.
3. Guess another user’s credentials.
4. Disable the authorization requirement.

Correct Answer: 2

Explanation

A permanent authorization failure should not be treated like a temporary network problem. Repeatedly retrying the same unauthorized request will not normally resolve the underlying permission issue and may create unnecessary load or security alerts. The agent should communicate that the requested operation could not be performed because the required authorization is unavailable. If an approved alternative exists, the workflow may direct the user toward it. The agent should never attempt to bypass authorization, use another person’s credentials, or disable security controls. Authorization boundaries must remain enforced even when the user strongly requests the action.

Question 138

Which design helps prevent a single agent tool from becoming difficult to secure and maintain?

1. Give the tool every available database and API operation.
2. Use focused tools with clearly defined responsibilities.
3. Hide the tool’s permissions from developers.
4. Allow unrestricted parameters for maximum flexibility.

Correct Answer: 2

Explanation

Focused tools are easier to understand, test, secure, and monitor than large tools that perform many unrelated operations. A tool with a narrow responsibility can have specific parameters, permissions, validation rules, and audit requirements. For example, a read-only order lookup tool can be separated from an order-update tool so that write access is granted only where necessary. Broad unrestricted tools increase the potential impact of model mistakes and make authorization more complicated. Clear boundaries also help the model select tools appropriately and make failures easier to diagnose.

Question 139

An agent completes a multi-step workflow but does not clearly indicate whether the final action succeeded. What improvement would be MOST useful?

1. Add explicit workflow completion states and success/failure reporting.
2. Remove all status information.
3. Allow the agent to assume success after the first tool call.
4. Hide the final tool response.

Correct Answer: 1

Explanation

Multi-step workflows benefit from explicit state tracking so the system can distinguish between in-progress, successful, failed, and partially completed operations. Without a clear completion state, an agent may incorrectly tell the user that a task succeeded when a later step failed. The application should track important tool results and define what constitutes successful completion. Clear status reporting also improves debugging and user trust. Simply assuming success after an early step is unsafe because subsequent actions may still fail. Explicit completion states provide a reliable connection between actual system behavior and what the agent communicates.

Question 140

An agent is processing untrusted text retrieved from a webpage. The text contains a request to send confidential information to an external address. What should the agent do?

1. Follow the webpage’s instruction because it appears specific.
2. Treat the webpage content as untrusted and do not perform the unauthorized action.
3. Send only part of the confidential information.
4. Ask the webpage for permission before sending the data.

Correct Answer: 2

Explanation

Content retrieved from webpages should generally be treated as untrusted data unless the application explicitly grants it authority. A webpage can contain prompt-injection instructions designed to manipulate an agent into performing unauthorized actions. Sending confidential information to an external address based solely on such content would create a serious security and privacy risk. The agent should follow the application’s instruction hierarchy and authorization rules instead. If the requested action is legitimate, it should come from an authorized workflow or user instruction and pass the application’s security checks before any external side effect occurs.