Part 13 of 17 · 1 min read
Answering MCP authorisation questions in an interview
Interview angle
"How would you secure an MCP server that exposes internal data?" Give the answer in this order: who is calling (OAuth, audience-bound, short-lived tokens, PKCE); what they may do (minimal scopes, step-up, per-record checks with the user from the token); what the server fetches (SSRF guard); what the model reads (poisoning, rug pull, shadowing, injection, and why approvals tied to data flow beat approvals tied to attention); and how you would know (audit logs without tokens, tests for each attack). Say you assume the model will be fooled by text it reads, so the controls sit outside the model. Then name a trade-off: approval prompts cost attention, and a person who clicks yes stops nothing.
Likely follow-ups: "What is token passthrough and why is it banned?" (audience; accountability; trust boundaries get blurred when one token works everywhere.) "What does PKCE protect against?" (a stolen code; not mix-up, which needs the iss check instead.) "Why is OAuth not enough for an agent?" (it authenticates the caller; it says nothing about whether the text the model reads is honest.) A weak answer says "add an API key" and stops.