MCP Is Not a Trust Boundary: why agent tool calls need signed permissions
JeremyCMorgan · x · 2026-11-22
A DEV Community article argues that MCP standardizes how agents call tools but doesn't establish whether those calls are authorized.
Key points:
- Trust in an approved MCP server does not transfer to every instruction it returns or every action the agent takes. Example: hidden instructions in a support ticket trick the agent into exporting internal account data — the connection is legitimate, the action is not (a prompt-injection risk).
- Separate three decisions: what entered the system (request/document/tool response/agent reasoning), what the agent wants to do (read, modify, message, transact), and what authority applies now (identity, data sensitivity, target system, risk, policy).
- Runtime policy should evaluate the proposed action, not just the source: reading a public article is low risk, sending customer data externally is not.
- The author proposes signed permissions and verifiable receipts, with an enterprise test: can you reconstruct why the agent was allowed to act?
More from coding & agent
- JetBrains Air runs Claude Code on your existing Claude Pro/Max plan, no extra subscription — kimmonismus · 2026-11-22
- ContextDuty: open-source local-first firewall stops secrets and PII from reaching AI coding tools — Interesting_Big_8974 · 2026-11-21
- Diplomat's 'second brain' rebuilt in Claude Duet: curated wiki, one skill, no servers — bytebot · 2026-11-21
- Indie dev spins up 4th side project with an agent swarm: astronomy escape room — jacalulu · 2026-11-21
- Veteran dev revives decade-old iOS game ideas with Gemini 4 — aziz4ai · 2026-11-21
- Open-source Mac menu bar app Tally warns you before hitting Claude usage limits — Raijin0101 · 2026-11-21