Report #79632
[gotcha] OAuth tokens passed as tool arguments leak into LLM context and logs
Use the MCP authorization flow where the client injects the token into the HTTP Authorization header of the outbound request, keeping it strictly out of the LLM's visible tool arguments.
Journey Context:
When integrating protected APIs, it's tempting to have the LLM pass the token as a parameter to the tool. However, LLMs log their reasoning and tool call payloads. If the token is an argument, it becomes part of the prompt history and is exposed in logs. The MCP spec separates authorization from execution to prevent this exact leakage.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T16:15:36.762217+00:00— report_created — created