Question 110
A tool catalog has separate search_web and internal_doc_search methods. The assistant uses the public search for confidential records. Which fix is strongest?
- Give the public search tool access to all private records
- Separate data scopes and restrict confidential retrieval
- Change every tool name to query without documenting boundaries
- Ask the model to memorize confidentiality rules without restrictions
Explanation
Correct Answer
B:Semantic descriptions and runtime permissions should align. Public search must not become a backdoor for internal data handling.
Incorrect Answers
A:Expanding external reach worsens confidentiality. It does not address inappropriate tool selection.
C:Generic names obscure where data is allowed to flow. Distinct scope information is necessary.
D:Prompts help but cannot guarantee safe data access. The executor must enforce allowed destinations.
Question 111
A data extractor occasionally returns prose even though downstream code expects tool arguments. How should the app tighten the interface?
- Accept prose and silently drop unparsed fields
- Require the extraction call and validate the returned schema
- Ask users to manually rewrite the response into JSON
- Parse any arbitrary prose with fragile regular expressions
Explanation
Correct Answer
B:Tool-choice controls can ensure the response enters a typed path. Runtime validators protect against semantically invalid data.
Incorrect Answers
A:Missing extraction fields may be material. A silent fallback hides data loss.
C:Manual conversion defeats the automation goal. The model and host should provide a programmatic interface.
D:Free-form text is more error-prone than a defined structure. The requirement calls for a stable machine-readable contract.
Question 112
A support agent should be able to invoke either lookup_order or lookup_return as appropriate, but no mutation tools. Which control combination works?
- Expose refund_order as well because all support tickets concern orders
- Expose only the two read tools and let auto choose between them
- Force lookup_order for every ticket including returns-only questions
- Disable the lookups and ask for a best-effort answer from memory
Explanation
Correct Answer
B:The model retains discretion between authorized queries while the capability set prevents writes. Clear descriptions make routing more reliable.
Incorrect Answers
A:A billing write is not necessary for lookup. Broader privileges create risk without aiding read operations.
C:A fixed tool does not fit both intents and may retrieve the wrong record type. This overconstrains choice.
D:Live order details cannot be reliably inferred. Read-only tools are the relevant sources.
Question 113
A security scanner is allowed to call a tool only after receiving a signed approval token. What additional control is required beyond tool_choice?
- Choose a more deterministic model and remove approval checks
- Put a screenshot of a token in the prompt and skip cryptographic checks
- The tool executor must verify the signed token before running the scan
- Set tool_choice to any and assume the forced call is authorized
Explanation
Correct Answer
C:Tool-choice constrains model-generated calls but does not establish authority. Execution checks must validate approval and scope.
Incorrect Answers
A:Sampling reliability does not replace cryptographic or policy controls. Authorization is an external system property.
B:Textual claims are not trusted authorization evidence. Signed tokens should be validated by controlled application code.
D:Requiring a tool use is not permission verification. The call could still violate business rules.
Question 114
A task-oriented agent has no reason to call tools for a general explanation but must be prevented from accessing an external network. What configuration expresses the boundary?
- Force any tool call even though no tool is required
- Exclude network-capable tools from the agent's permitted capabilities
- Let a child agent inherit all network tools and hide the result
- Expose the browser tool and ask Claude to avoid it
Explanation
Correct Answer
B:Removing outbound-access tools blocks that path regardless of model intent. The agent can still provide a text explanation from allowed context.
Incorrect Answers
A:Mandatory tool use adds an unnecessary action and can pick an irrelevant capability.
C:Subagent inheritance can bypass the intended isolation if not scoped. Network access must be restricted at the executor and tool layer.
D:A prompt-only boundary is weaker than capability restriction. Accidental tool selection remains possible.
Question 115
A development team wants Claude Code to use a shared internal knowledge server with MCP. Which integration concept is correct?
- Configure an MCP server connection that exposes discoverable tools or resources
- Assume installing an MCP server automatically grants all team members data access
- Paste a raw API token into every model prompt instead of configuring MCP
- Train a new language model on the knowledge base whenever documents change
Explanation
Correct Answer
A:MCP provides a standardized client-server interface for capabilities and data. Claude Code must be able to discover the published functionality to use it.
Incorrect Answers
B:Connectivity does not remove backend authorization. Each caller remains subject to the server's access controls.
C:Tokens in prompts are insecure and do not define a structured server interface. Authentication and tool discovery should be handled by the integration.
D:MCP supports live access to changing information. Model training is unnecessary for this integration requirement.
Question 116
A team needs an MCP connection shared through repository configuration so new developers get the same tool definitions. Which scope is suitable?
- Use approved project-scoped configuration for team access
- Configure the server only in one developer's personal user scope
- Publish production credentials inside the shared config
- Change an unrelated global model setting instead of configuring the server
Explanation
Correct Answer
A:Project scope distributes a consistent integration with the repository. Secrets must remain outside committed configuration.
Incorrect Answers
B:User scope is useful for private preferences but does not establish a shared project default for the team.
C:Shared connection metadata should not embed secrets. Credentials require secure injection or user-specific authorization.
D:Model preferences do not declare MCP server endpoints and capabilities. The integration needs an MCP configuration.
Question 117
An engineer needs an MCP server only for personal experimentation and does not want colleagues to inherit it. Which scope should be used?
- Rename the server to private without changing its scope
- Add the server to every shared workspace independently
- Commit the experimental server as a project requirement
- Use user-scoped configuration for the experimental server
Explanation
Correct Answer
D:A user-specific connection remains separate from repository defaults. It avoids unexpectedly adding the experiment to teammates' tool lists.
Incorrect Answers
A:A label does not alter which configurations are loaded. Visibility follows the selected scope.
B:Repetition increases maintenance and still risks team exposure. The personal scope is more direct.
C:Project configuration would distribute the dependency to the team. That conflicts with the stated personal-only intent.
Question 118
A local MCP integration launches a trusted helper process through standard input and output. Which transport pattern fits?
- Use a browser URL as the only command field
- Write tool outputs into CLAUDE.md instead of using a transport
- Use an stdio MCP server configured with the executable command
- Require an arbitrary public web endpoint for every local helper
Explanation
Correct Answer
C:Stdio is a common local process transport. The client communicates with the launched process through standard streams.
Incorrect Answers
A:A local process needs an executable and arguments. A URL alone does not launch the stdio helper.
B:Project instructions cannot act as a live MCP request-response channel. Runtime communication still needs a supported transport.
D:A locally launched stdio server does not need Internet exposure. Adding a public endpoint increases complexity and risk.
Question 119
A remote service exposes MCP over an HTTP endpoint. What should the client configuration primarily describe?
- A local shell command that assumes the server runs on the developer's machine
- A model fine-tuning job for the server's entire API surface
- The supported remote transport and endpoint plus appropriate authentication handling
- Only a prompt telling Claude that the service exists
Explanation
Correct Answer
C:Remote MCP requires a connection description that matches the server. Authentication remains a separate security concern from tool discovery.
Incorrect Answers
A:Local process launch is a different transport pattern. It will not connect to an already hosted remote endpoint.
B:MCP provides live interfaces. Training a model does not replace network configuration and tool execution.
D:A declaration in prose does not establish a client connection or expose server capabilities.
Question 120
A team connects Claude Code to an internal MCP document service for collaborative research. Which THREE practices support a reliable and secure integration? Choose THREE.
- Paste confidential service credentials into the model's visible prompt
- Preserve structured tool errors so Claude can recognize failed retrievals
- Train a new model every time the MCP resource catalog changes
- Configure an appropriate client transport and reachable MCP endpoint
- Limit exposed capabilities and enforce user permissions on the service
- Assume a connected server makes every private resource public by default
Explanation
Correct Answers
B:Transport success does not imply business success. Actionable error messages support retries, clarification and honest uncertainty.
D:The client must establish an actual server connection before advertised tools or resources can be discovered. Prompt text alone is not connectivity.
E:MCP standardization does not replace authorization. The server and client should prevent unrelated destructive actions for a read-only role.
Incorrect Answers
A:Secrets should be managed by the runtime and authentication layer. Exposing them as prompt context creates unnecessary disclosure risk.
C:MCP advertises current capabilities and data dynamically. Model retraining is not the mechanism for refreshing client discovery.
F:Connection and entitlement are different properties. The backend still controls what a user may access.