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

 

Question 221

An agent must process a customer’s request containing sensitive personal information. What is the most appropriate approach?

1. Share all customer information with every downstream service
2. Minimize the sensitive data sent to each component to what is necessary
3. Store the information permanently in the prompt
4. Include the information in every application log

Correct Answer: 2

Explanation

Sensitive information should be handled according to data-minimization principles. Each component should receive only the information necessary to perform its specific task. Sending unnecessary personal data to downstream services increases privacy and security exposure and may create additional compliance requirements. Similarly, permanently storing sensitive information in prompts or logging it indiscriminately creates unnecessary risk. The application should use appropriate access controls, encryption, retention policies, and redaction where needed. Minimizing data does not mean removing information required for legitimate processing; it means avoiding unnecessary collection, transmission, storage, and exposure.

Question 222

A tool is responsible for changing a user’s account email address. Which design provides the clearest security boundary?

1. A general-purpose database administration tool
2. A tool that can modify every account field
3. A tool with unrestricted write access
4. A narrowly scoped email-update tool with authorization checks

Correct Answer: 4

Explanation

A narrowly scoped tool creates a clearer security boundary because its purpose and permissions are limited to the specific operation required. An email-update tool can validate the account identifier and new email address while enforcing authorization before making the change. A general database tool or unrestricted account-update tool gives the agent unnecessary capabilities that could be misused or triggered accidentally. Narrow tools also simplify testing, monitoring, and auditing because developers can clearly identify what each operation is allowed to do. Sensitive account changes should additionally use appropriate identity verification and confirmation controls.

Question 223

An agent receives a tool error indicating that the request is permanently invalid because a required identifier is missing. What should it generally do?

1. Retry the same request indefinitely
2. Fabricate the missing identifier
3. Correctly obtain the missing information before retrying
4. Report that the operation succeeded

Correct Answer: 3

Explanation

A permanent validation error generally cannot be resolved by repeatedly sending the same invalid request. If a required identifier is missing, the agent should obtain the correct identifier through an authorized retrieval method or ask the user for it. Fabricating an identifier can cause the wrong record to be accessed or modified. Unlimited retries waste resources and may trigger rate limits without changing the outcome. The application should distinguish permanent input errors from temporary service failures so that retry behavior is appropriate to the error type. Clear structured errors make this distinction easier.

Question 224

Why is it useful to include tool descriptions that clearly state when a tool should and should not be used?

1. It helps the agent select the appropriate capability and avoid misuse
2. It guarantees that every tool call will succeed
3. It eliminates backend authorization
4. It prevents tools from returning errors

Correct Answer: 1

Explanation

Clear tool descriptions provide important guidance about a tool’s purpose, required inputs, limitations, and appropriate use cases. This helps the agent distinguish between similar capabilities and reduces the chance of selecting a tool for an unintended operation. For example, a read-only customer lookup tool should clearly differ from a customer-update tool. Good descriptions do not replace backend security controls, input validation, or authorization. Instead, they complement those controls by making the available capabilities easier to understand and use correctly. Focused tool responsibilities further improve reliable tool selection.

Question 225

An agent must perform a financial operation that cannot easily be reversed. Which additional safeguard may be appropriate?

1. Removing transaction records
2. Requiring explicit confirmation or human review before execution
3. Allowing the model to execute automatically without checks
4. Giving the model broader financial permissions

Correct Answer: 2

Explanation

Irreversible or high-impact financial operations deserve stronger safeguards because mistakes can have significant consequences. Depending on the application’s risk model, the system may require explicit user confirmation, stronger authentication, human review, transaction limits, or multiple authorization steps before execution. The agent should also use validated parameters and appropriate audit logging. Giving the model broader permissions would increase risk rather than reduce it. Removing transaction records would make investigations more difficult. The exact control depends on the operation, but additional verification is generally appropriate when an action can cause substantial or difficult-to-reverse financial consequences.

Question 226

Which practice helps an agent avoid acting on stale information?

1. Never refresh external data
2. Use cached information indefinitely
3. Consider data freshness requirements and retrieve current information when necessary
4. Assume every stored value is current

Correct Answer: 3

Explanation

Some tasks depend on information that changes frequently, such as account balances, inventory levels, appointments, or service availability. In these cases, relying on old cached information can cause incorrect decisions or actions. The system should define appropriate freshness requirements and retrieve current data when the task requires it. Caching can still be useful when the information changes infrequently or when a defined freshness window is acceptable. The key is to understand the task’s requirements rather than assuming that all stored information remains accurate indefinitely.

Question 227

An application uses an agent to answer questions from a private knowledge base. What is important when retrieving documents for a specific customer?

1. Retrieve only documents the authenticated user is authorized to access
2. Retrieve every document and let the model decide what to reveal
3. Ignore document-level permissions
4. Use the customer’s name as the only security check

Correct Answer: 1

Explanation

Authorization should be enforced before private documents are provided to the agent. The retrieval system should ensure that the authenticated user or tenant is permitted to access each document. Relying on the model to decide which private information it should reveal is unsafe because models can make mistakes or be influenced by malicious content. A customer’s name alone is also not a sufficient security control. Retrieval systems should apply access-control filters at the data layer and maintain appropriate isolation between customers or tenants. This prevents unauthorized documents from entering the model’s context in the first place.

Question 228

An agent is processing a long conversation and approaching its context limit. What is a reasonable strategy?

1. Add unrelated information to increase context usage
2. Remove all system instructions
3. Repeat the entire conversation several times
4. Summarize or retrieve only the information necessary for the next step

Correct Answer: 4

Explanation

When context becomes large, the application can reduce unnecessary information by summarizing earlier conversation content or retrieving only the portions relevant to the current task. Important instructions, constraints, decisions, and required facts should be preserved. Simply removing system instructions can undermine application behavior, while repeating the entire conversation wastes context capacity. Adding unrelated information provides no benefit and can make relevant information harder to identify. Effective context management helps maintain performance, reduce token usage, and ensure that the model has the information necessary to complete the current task accurately.

Question 229

What is the main purpose of using provenance metadata with retrieved information?

1. To make the model’s response longer
2. To identify where information came from and support verification
3. To remove the need for retrieval
4. To guarantee that the source is always correct

Correct Answer: 2

Explanation

Provenance metadata helps identify the origin of information used by an agent. It can include source identifiers, document names, timestamps, URLs, database records, or other information that allows the application or user to understand where a statement came from. Provenance is particularly useful when sources conflict, information needs verification, or an audit trail is required. It does not guarantee that the underlying source is accurate. Instead, it provides context that supports validation and accountability. Keeping useful provenance can also help developers diagnose retrieval problems and understand why particular information appeared in an agent’s context.

Question 230

An agent needs to update a customer’s address, but the user provides an incomplete address. What should the agent do?

1. Guess the missing information
2. Use an unrelated address from another customer
3. Ask for the missing information required to complete the update
4. Submit the incomplete address without validation

Correct Answer: 3

Explanation

An address update should not be performed using guessed or incomplete information when required fields are missing. The agent should identify which information is necessary and ask the user for those details before executing the change. This is especially important for persistent customer records because an incorrect update can affect shipping, billing, identity verification, or other processes. Backend validation should also check the submitted values before the update occurs. Using another customer’s address would be a serious privacy and data-integrity issue. Asking a focused clarification question is the appropriate response when required information is unavailable.

Question 231

Which approach is most appropriate for handling an agent’s access to production credentials?

1. Expose credentials directly in the model’s prompt
2. Store credentials in secure application infrastructure and restrict access
3. Put credentials in user-visible tool descriptions
4. Include credentials in every log entry

Correct Answer: 2

Explanation

Production credentials should be managed by secure application infrastructure rather than exposed directly to the model or users. Secret-management systems, environment-level controls, short-lived credentials, and narrowly scoped permissions can reduce the risk of credential exposure. Credentials should not be placed in prompts, tool descriptions, or ordinary logs because these surfaces may be accessible to unintended parties or retained longer than necessary. The agent can request an authorized operation through a controlled tool while the underlying application handles credentials securely. This separation helps prevent the model from directly handling sensitive authentication secrets.

Question 232

An agent receives a request that could be interpreted in two substantially different ways, both leading to different external actions. What should it generally do?

1. Choose whichever interpretation is easiest
2. Perform both actions automatically
3. Guess the user’s intended action
4. Ask a clarification question before taking the external action

Correct Answer: 4

Explanation

When ambiguity could result in materially different external actions, the agent should clarify the user’s intent before creating a side effect. For example, if “cancel the booking” could refer to either of two reservations, guessing could cancel the wrong reservation. Asking a focused clarification question prevents avoidable errors and gives the user control over the final action. This is especially important for financial transactions, account changes, communications, deletions, and other irreversible or high-impact operations. For low-risk tasks, the application may sometimes use reasonable defaults, but significant ambiguity should generally be resolved explicitly.

Question 233

Why should an agent’s execution budget be bounded?

1. To prevent uncontrolled loops and excessive resource consumption
2. To guarantee perfect answers
3. To eliminate the need for error handling
4. To ensure every task uses the maximum number of tool calls

Correct Answer: 1

Explanation

An execution budget limits resources such as tool calls, processing time, tokens, or retries that an agent can consume during a task. Without appropriate limits, an agent could become stuck in a loop, repeatedly call failing tools, or consume excessive infrastructure resources. Budgets help make system behavior more predictable and protect against accidental or malicious workloads. They do not guarantee correctness, so applications still need validation, error handling, and appropriate termination conditions. When a budget is reached, the system should communicate the incomplete state clearly rather than pretending that the task was successfully completed.

Question 234

A tool returns a structured response containing a status field of partial. What should the agent understand from this result?

1. The entire operation definitely succeeded
2. The tool should always be called again indefinitely
3. Some requested work completed, but additional work remains or failed
4. The response should be ignored

Correct Answer: 3

Explanation

A partial status indicates that the requested operation was not fully completed. The agent should inspect the structured response to determine which parts succeeded and which remain incomplete or failed. Depending on the workflow, it may be appropriate to retry a failed portion, ask the user for additional information, or clearly report the partial result. Treating a partial response as complete success can cause the user or downstream system to assume that all requested work was performed. Structured completion states therefore help agents communicate accurate workflow status and choose appropriate next steps.

Question 235

Which approach is most appropriate when an external API returns sensitive customer information that the agent does not need?

1. Pass all fields to every downstream component
2. Remove or avoid unnecessary sensitive fields before further processing
3. Store the sensitive response permanently
4. Include all fields in application logs

Correct Answer: 2

Explanation

If the agent does not need particular sensitive fields, those fields should ideally be excluded before the data moves to another component. This reduces unnecessary exposure and follows data-minimization principles. For example, an order-status task may require an order number and shipping status but not a customer’s full payment details. Passing unnecessary information through the system increases privacy and security risks and may complicate compliance obligations. Sensitive data should also be protected through appropriate access controls, retention policies, and logging practices. Minimization is therefore an important part of secure agent architecture.

Question 236

An application wants to determine whether a model update changed the agent’s tool-selection behavior. What should developers compare?

1. Only the number of words in final responses
2. Only the application’s startup time
3. Tool-call traces and evaluation results from before and after the update
4. The model’s name only

Correct Answer: 3

Explanation

Tool-call traces provide useful evidence about which tools the agent selected, what parameters it supplied, and how it responded to tool results. Comparing traces before and after a model update can reveal changes that may not be obvious from final text alone. A regression suite can further test whether the new model continues to select appropriate tools across representative scenarios. Response length or model naming does not provide meaningful evidence about tool-selection behavior. Versioned traces and repeatable evaluations therefore provide a stronger basis for identifying behavioral changes after model or configuration updates.

Question 237

What should happen if a user requests an operation for which the agent does not have authorization?

1. Execute it anyway because the user requested it
2. Ask the model to bypass the permission system
3. Report that the operation cannot be performed under the current authorization
4. Use another user’s credentials

Correct Answer: 3

Explanation

Authorization boundaries should be enforced regardless of what the user asks the agent to do. If the authenticated identity lacks permission for an operation, the system should deny the action and communicate the limitation appropriately. The agent should never attempt to bypass access controls, use another person’s credentials, or manipulate the authorization system. Backend authorization is particularly important because model instructions alone are not a reliable security boundary. Depending on the application, the user may be directed toward an approved process for requesting additional permissions, but the unauthorized operation itself should not be executed.

Question 238

An agent needs to send a calendar invitation to several participants. Which information should be validated before sending?

1. Only the agent’s preferred formatting
2. The event details, participants, time, and relevant time zone
3. The number of words in the invitation
4. Unrelated customer records

Correct Answer: 2

Explanation

Before sending a calendar invitation, the application should validate the information that determines the actual event and its recipients. This includes participant addresses, event title or purpose, date, start and end times, and the applicable time zone. Missing or ambiguous information can result in invitations being sent to the wrong people or scheduled at an unintended time. Because sending invitations creates an external side effect, the system may also require confirmation depending on the workflow. Validating relevant fields helps prevent avoidable scheduling errors while keeping unrelated data outside the operation.

Question 239

Why is it useful to distinguish temporary errors from permanent errors in an agent workflow?

1. It allows the system to choose appropriate retry or recovery behavior
2. It guarantees every operation will eventually succeed
3. It removes the need for structured errors
4. It causes the agent to retry every request indefinitely

Correct Answer: 1

Explanation

Different errors require different recovery strategies. A temporary service outage or rate-limit response may justify a bounded retry with backoff, while an invalid identifier or unauthorized request generally requires correction or user intervention rather than another identical attempt. Distinguishing these categories helps prevent unnecessary retries, reduces resource consumption, and improves user communication. Structured error codes and clear tool responses can make this classification easier for the application to implement. The distinction does not guarantee success; it simply allows the system to respond more intelligently to different failure conditions.

Question 240

An AI agent is deployed across multiple customer tenants. Which architecture principle is especially important?

1. Allow the model to select any tenant’s data
2. Share all retrieved records across tenants
3. Enforce tenant isolation at the application and data-access layers
4. Rely only on user-provided tenant names

Correct Answer: 3

Explanation

Tenant isolation is essential in multi-tenant systems because one customer’s data must not become accessible to another customer. The application should enforce tenant boundaries through authenticated identity, authorization checks, database-level filtering, and other appropriate controls. Relying on the model to select the correct tenant or trusting a user-provided tenant name is insufficient because either can be incorrect or manipulated. Retrieval systems and tools should receive the appropriate tenant context and enforce it independently of model behavior. Strong isolation reduces the risk of accidental cross-customer data exposure and supports safer agent operation across shared infrastructure.