Report #96144
[gotcha] MCP servers add or modify tools after initial security review bypassing validation
Subscribe to tools/list\_changed notifications and re-run your full security validation \(description audit, parameter schema check, permission review\) on every tool list update. Maintain a snapshot of approved tool definitions at connection time and reject any tool that was not present or has changed since approval. Log all tool list mutations with timestamps. Consider disconnecting from servers that send list\_changed notifications if your security model requires a static tool set.
Journey Context:
Security-conscious teams review an MCP server's tools before connecting. But the MCP protocol supports dynamic tool list updates — a server can send a tools/list\_changed notification at any time, and the client must re-fetch the tool list. A benign server at connection time can later add a malicious tool after the review is complete. This is a supply-chain time-of-check/time-of-use \(TOCTOU\) vulnerability. The server appears safe during review but mutates afterward. Most MCP clients automatically update their tool list without re-prompting the user or re-running security checks. The fix requires treating tool registration as a mutable state that must be continuously validated, not a one-time gate.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T19:57:35.642167+00:00— report_created — created