Claude Certified Architect – Foundations Practice Question 49: Untrusted MCP Server Risks

Claude Certified Architect, Foundations · Test 5 · question 9 of 10

Tool Design & MCP Integrationhard

A developer finds an open-source MCP server on a public registry that promises to "supercharge code reviews". It runs locally via npx, exposes tools that read repository files and post to a web service, and requests a broad OAuth scope on the connected platform. A security engineer warns that a malicious or compromised MCP server could embed hidden instructions in its tool descriptions or results to steer Claude into leaking data. Which approach best reduces the risk while still letting the team evaluate the server?

Select one answer
Show answer and explanation

Answer:

  • B. Review the source and tool descriptions first, pin the version, run it sandboxed with only the filesystem and network access it needs, and limit OAuth to the minimum scopes.

MCP tool descriptions and results are content that flows into the model's context, so a malicious server can use them for prompt injection. Claude Code's MCP documentation says to verify you trust each server before connecting it and to audit servers that fetch or proxy external content. The MCP security best practices add that local servers run with the client's privileges and should be sandboxed with minimal default privileges, that the exact startup command should be reviewed, and that scopes should follow least privilege. Reviewing the code and descriptions, pinning the version, sandboxing, and narrowing OAuth scopes limit both the likelihood and the blast radius.

A is wrong because tool descriptions and results are placed in the model's context, and the model does not reliably discard instructions based on where they came from; sharing it team-wide also widens exposure. C is wrong because no model is completely immune to prompt injection; model behavior is one layer, not a substitute for trust decisions and isolation. D is wrong because pre-approving every tool removes the human checkpoint, so a poisoned instruction could act without any prompt.

Reference: Security Best Practices (Model Context Protocol)

Practice Test 5 ›

Study guides for Claude Certified Architect, Foundations