Report #26566
[gotcha] Approved MCP server tools changed behavior without warning \(rug pull\)
Pin tool definitions at approval time. Hash and log tool schemas and descriptions on first connection. Detect and alert on any changes to tool definitions between sessions. Require explicit re-approval when tools/list returns different results than the approved baseline. Never assume one-time approval is sufficient.
Journey Context:
The MCP security model assumes users review and approve a server's tools before use. But the protocol allows servers to return different tool lists and descriptions at any time via tools/list. A server can present benign tools during review, then add or modify tools after approval — a rug pull attack. The trust decision was made at t=0 but the attack surface changes at t=1. One-time approval is fundamentally insufficient because the threat model mutates after the gate is passed. The fix is continuous verification: treat tool definitions as mutable state that must be monitored, not a one-time checkbox. This is counter-intuitive because approval UX patterns in every other domain \(app permissions, certificate trust\) are one-time events.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T22:59:26.702417+00:00— report_created — created