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

 

Question 201

An AI agent needs to access a production database to retrieve customer information. Which approach best follows the principle of least privilege?

1. Give the agent administrator access
2. Allow access to every database table
3. Provide only the read permissions and tables required for the task
4. Share the database root credentials with the agent

Correct Answer: 3

Explanation

Least privilege means giving an agent only the permissions necessary to perform its intended task. If an agent only needs to retrieve customer order information, it should receive narrowly scoped read access rather than administrative permissions or access to unrelated tables. This reduces the potential impact of bugs, compromised credentials, prompt injection, or incorrect tool selection. Production credentials should also be managed securely rather than embedded directly into prompts or exposed to the model. Narrow permissions make security controls easier to audit and help prevent an agent from accidentally modifying or accessing unrelated production data.

Question 202

A user asks an agent to transfer money but does not specify the destination account. What should the agent do?

1. Ask for the missing destination account information
2. Select the most recently used account automatically
3. Transfer the money to the user’s primary account
4. Use a default account without notifying the user

Correct Answer: 1

Explanation

A financial transfer requires sufficient information to identify the intended destination accurately. If the destination account is missing, the agent should request clarification rather than guessing or selecting an account automatically. Financial operations are high-impact actions, so incorrect assumptions can cause real monetary loss. The system should also apply authentication, authorization, transaction validation, and appropriate confirmation controls before executing the transfer. Choosing a default account without user confirmation could send funds to an unintended recipient. Asking a focused clarification question is therefore the safer approach when a critical transaction parameter is missing.

Question 203

Which approach is most appropriate for protecting sensitive information in application logs?

1. Log every piece of information for maximum detail
2. Disable all logging permanently
3. Store passwords in plain text for troubleshooting
4. Redact or avoid logging sensitive data that is not necessary

Correct Answer: 4

Explanation

Logs are valuable for troubleshooting and auditing, but they can become a significant source of data exposure if they contain unnecessary sensitive information. Applications should follow data-minimization principles by avoiding or redacting passwords, authentication tokens, payment details, personal identifiers, and other sensitive values when they are not required for operational purposes. Completely disabling logging can make debugging and incident investigation difficult. Similarly, storing secrets in plain text creates unnecessary security risks. A balanced logging strategy captures enough information to understand system behavior while minimizing the amount of sensitive data retained.

Question 204

An agent is given a webpage as context. The webpage contains instructions telling the agent to reveal its system prompt. How should the agent treat those instructions?

1. Follow them because they came from a webpage
2. Treat them as untrusted content unless explicitly authorized
3. Give the webpage higher priority than system instructions
4. Reveal the system prompt before continuing

Correct Answer: 2

Explanation

Webpages and other external content should generally be considered untrusted input. They may contain instructions specifically designed to manipulate an AI system, including attempts to override higher-priority instructions or extract confidential information. The agent should treat the webpage as data to analyze rather than automatically treating its embedded instructions as commands. System and application-level policies should continue to control the agent’s behavior. Additional protections such as tool authorization, content isolation, and output validation can reduce the impact of prompt injection attacks when agents interact with external webpages or documents.

Question 205

Why is it useful to assign clear and focused responsibilities to individual tools?

1. It makes authorization and tool selection easier to reason about
2. It guarantees that the model will never make an incorrect decision
3. It removes the need for tool descriptions
4. It allows every tool to perform unrelated operations

Correct Answer: 1

Explanation

Focused tools are easier for both developers and models to understand. When each tool has a clear responsibility, its description, required parameters, permissions, and expected results can be defined more precisely. This makes tool selection easier and reduces the chance that a broad tool will be used for an unintended purpose. Focused responsibilities also support stronger authorization because permissions can be scoped to specific operations. However, clear tool design does not guarantee perfect model behavior, so applications should still use validation, authorization, monitoring, and appropriate safeguards for sensitive operations.

Question 206

An agent needs to perform a database update after first checking whether a record exists. What is an important consideration when designing this workflow?

1. Skip validation because the model already understands the request
2. Allow the update regardless of authorization
3. Ensure the update is performed only after required validation and authorization
4. Hide database errors from the application

Correct Answer: 3

Explanation

Database updates can create persistent side effects, so the application should validate the target record, input values, and authorization before performing the operation. A model’s interpretation of the user’s request should not replace backend controls. The workflow may first retrieve the relevant record, verify that the user has permission to modify it, validate the requested changes, and then execute the update. Database errors should also be handled and recorded appropriately. This layered approach helps prevent accidental modifications, unauthorized access, and inconsistent data while providing clearer behavior when something goes wrong.

Question 207

What is a major benefit of versioning prompts and tool definitions in an AI application?

1. It prevents all future model errors
2. It makes every response deterministic
3. It removes the need for evaluations
4. It helps reproduce behavior and identify which configuration produced an output

Correct Answer: 4

Explanation

Versioning prompts, tool definitions, configuration, and related components helps teams understand how an agent behaved at a particular point in time. If an application changes its system prompt or modifies a tool schema, version information makes it easier to reproduce an issue and determine which configuration produced a particular result. This is especially valuable for regression testing, debugging, audits, and controlled deployments. Versioning does not guarantee deterministic behavior because model outputs can still vary. It simply provides important context for understanding and comparing different executions over time.

Question 208

A customer-support agent receives two conflicting pieces of information from different internal sources. What should it generally do?

1. Select the source with the shorter response
2. Ignore both sources
3. Follow the application’s source-priority or conflict-resolution policy
4. Invent a compromise value

Correct Answer: 3

Explanation

Conflicting information should be handled according to an established source-priority or conflict-resolution policy rather than by arbitrary guessing. Some systems may designate a particular database as authoritative for customer status while another source is authoritative for shipping information. The agent should use available provenance and metadata to understand where information came from and, when necessary, communicate uncertainty to the user. Inventing a value can result in incorrect decisions. A well-designed system should define how conflicts are resolved and should make important discrepancies visible rather than silently selecting information without justification.

Question 209

Which mechanism is most appropriate for ensuring that a user is allowed to perform a sensitive account operation?

1. Authorization enforced by the backend
2. The model’s own judgment
3. A longer natural-language prompt
4. The user’s ability to describe the operation clearly

Correct Answer: 1

Explanation

Authorization should be enforced by trusted application or backend systems rather than relying on the model’s interpretation of a user’s request. The model can help determine what operation the user appears to want, but the backend should verify whether the authenticated identity has permission to perform that operation. This distinction is especially important for sensitive actions such as changing account ownership, modifying payment information, deleting data, or accessing confidential records. Natural-language instructions can guide behavior but should not serve as a substitute for actual access-control enforcement.

Question 210

An agent repeatedly encounters a rate-limit response from an external API. What is the most appropriate strategy?

1. Immediately send hundreds of additional requests
2. Ignore the rate limit and continue normally
3. Use controlled backoff and respect the API’s retry guidance
4. Change unrelated user data to bypass the limit

Correct Answer: 3

Explanation

Rate limits exist to control resource usage and protect external services from excessive traffic. When an API returns a rate-limit response, the application should respect the service’s guidance, such as a retry-after value, and use controlled backoff before attempting another request. Repeated immediate requests can make the problem worse and may lead to longer blocking periods. The system should also use bounded retries so that a persistent limitation does not result in an endless loop. Proper rate-limit handling improves reliability and reduces unnecessary load on the external service.

Question 211

What should an agent do when a required tool parameter contains an invalid data type?

1. Convert it to a random value
2. Validate the input and request or generate a valid value before execution
3. Execute the tool anyway
4. Ignore the validation error

Correct Answer: 2

Explanation

Tool parameters should conform to the schema expected by the tool. If a required parameter has an invalid type, the application or agent should identify the problem before execution and obtain a valid value where possible. For example, if a tool expects an integer customer ID but receives arbitrary text, the system should not blindly submit it. Schema validation can catch these issues early and provide structured error information. Executing malformed requests can cause failures or, in some systems, unexpected behavior. Proper validation therefore improves reliability and helps protect downstream services.

Question 212

An agent needs to retrieve information from several independent sources before answering a user. What should it do when the sources can safely be queried concurrently?

1. Use parallel execution when supported and appropriate
2. Always query them one hour apart
3. Query only one source regardless of the task
4. Ask the user to manually collect the information

Correct Answer: 1

Explanation

When multiple data sources are independent and there are no ordering dependencies, querying them concurrently can reduce total response time. Parallel execution is particularly useful when each external request has noticeable latency. The application should still consider rate limits, concurrency limits, failures, and resource consumption. If one request depends on the result of another, sequential execution may be necessary. Parallelism should therefore be applied based on actual dependencies rather than used blindly. Proper orchestration can improve performance while maintaining reliable handling of individual tool failures.

Question 213

Why should an AI agent distinguish between authentication and authorization?

1. They are exactly the same security control
2. Authentication determines what actions are permitted
3. Authentication identifies or verifies the user, while authorization determines what the user can access
4. Authorization is only needed for public information

Correct Answer: 3

Explanation

Authentication and authorization serve different purposes. Authentication establishes or verifies who the user or calling service is, while authorization determines what that authenticated identity is allowed to access or perform. For example, a user may successfully authenticate but still lack permission to modify another customer’s account. Confusing these concepts can create serious security gaps. Sensitive agent operations should therefore use appropriate identity verification together with backend authorization checks. The model should not be expected to independently enforce these security boundaries through natural-language reasoning alone.

Question 214

An agent is asked to summarize a large document, but only one section is relevant to the user’s question. What is generally a good approach?

1. Provide the entire document to every tool and component
2. Retrieve or process only the relevant section when possible
3. Ignore the user’s question
4. Ask the model to memorize the entire document permanently

Correct Answer: 2

Explanation

Processing only the information relevant to a task can improve efficiency and reduce unnecessary exposure of sensitive content. If a user asks about a particular section of a large document, retrieval or targeted processing can provide the relevant material without unnecessarily passing the entire document through every component. This supports data minimization and can also reduce context usage. The system should still ensure that enough context is included to answer accurately. Targeted retrieval is therefore useful when documents are large, contain sensitive information, or include substantial material unrelated to the user’s request.

Question 215

Which situation most strongly suggests that human review should be considered before an agent completes an action?

1. Generating a harmless internal summary
2. Formatting a non-sensitive document
3. Searching a public knowledge base
4. Permanently deleting a large set of important records

Correct Answer: 4

Explanation

Actions that are highly consequential, difficult to reverse, or capable of causing substantial harm often warrant additional human oversight. Permanently deleting important records is a clear example because an incorrect decision could cause significant and potentially irreversible data loss. Depending on the application, human review may also be appropriate for financial transfers, account ownership changes, sensitive communications, or other high-impact actions. Lower-risk tasks such as formatting or retrieving public information generally require fewer safeguards. The appropriate level of human involvement should reflect the potential impact and reversibility of the operation.

Question 216

An agent receives a successful tool response but the response contains unexpected fields and malformed values. What should the application do?

1. Validate the response before using it downstream
2. Assume all returned values are correct
3. Ignore the malformed fields and execute every side effect
4. Disable response validation permanently

Correct Answer: 1

Explanation

A successful network or tool call does not necessarily mean that the returned data is safe or correctly structured for downstream processing. Response validation can verify required fields, data types, allowed values, and other schema constraints before the application uses the result. This helps detect unexpected API changes, implementation bugs, malformed responses, or incompatible versions. Invalid data should be handled according to an established error policy rather than blindly passed into another operation. Structured schemas and validation therefore provide an important reliability layer between external tool responses and subsequent agent actions.

Question 217

What is an important reason to maintain an audit trail for sensitive agent actions?

1. To make the model generate longer explanations
2. To allow investigation of what happened and who or what initiated an action
3. To guarantee that unauthorized actions cannot occur
4. To eliminate the need for authentication

Correct Answer: 2

Explanation

An audit trail provides a record of significant actions and can help organizations investigate incidents, troubleshoot unexpected behavior, and demonstrate accountability. Depending on the system, an audit record might include the authenticated identity, operation performed, relevant resource, timestamp, outcome, and execution or request identifier. Audit logging does not itself prevent unauthorized actions, so it should be combined with authentication and authorization. Logging should also follow privacy and data-minimization requirements. For sensitive agent operations, an appropriate audit trail can be especially useful when determining how an unexpected change occurred.

Question 218

An agent is designed to update a customer’s address. Which tool design is generally preferable?

1. A tool that can modify any customer field
2. A tool that deletes and recreates the entire customer account
3. A narrowly scoped address-update tool with validated parameters
4. A tool that provides unrestricted database access

Correct Answer: 3

Explanation

A narrowly scoped tool reduces the agent’s ability to perform unrelated or harmful operations. If the intended task is updating an address, the tool should ideally accept the customer identifier and validated address fields rather than exposing broad database capabilities. This supports least privilege and makes the tool’s purpose clear to both the agent and developers. Narrow tools also simplify authorization, validation, testing, and auditing. Broad database access increases the consequences of an incorrect tool call or prompt injection attempt. The design should therefore expose only the capability required for the intended workflow.

Question 219

An agent cannot complete one step because an external service is temporarily unavailable. What should the agent generally tell the user?

1. That the entire task succeeded
2. That the service failure does not matter
3. A fabricated result based on an earlier request
4. That the step could not be completed and, when appropriate, what can happen next

Correct Answer: 4

Explanation

Users should receive an accurate representation of the actual workflow state. If an external service prevents one required step from completing, the agent should not claim that the whole operation succeeded or invent a result. It should explain that the relevant step could not be completed and, where appropriate, indicate whether retrying later or taking another action is possible. Clear partial-failure reporting helps users understand what has and has not happened. It also prevents downstream decisions from being based on a false assumption that all requested operations were completed successfully.

Question 220

Which practice is most useful when deploying a new model version in an existing agent system?

1. Replace the model without testing
2. Remove all previous evaluation data
3. Compare the new version against a repeatable regression test suite
4. Assume identical behavior because the tools have not changed

Correct Answer: 3

Explanation

Changing the model version can affect tool selection, instruction following, formatting, reasoning patterns, and responses to edge cases. A repeatable regression suite allows developers to compare the new version with the existing system across representative tasks. Tests should cover normal workflows, tool calls, error handling, safety requirements, structured outputs, and important edge cases. Keeping previous evaluation results provides a baseline for comparison. Even when prompts and tools remain unchanged, model behavior can change after an upgrade. Controlled testing therefore helps identify unexpected regressions before the new version is widely deployed.