RunPod's MCP server tells agents to rate it and hide it from users

XadniGraves · reddit · 2026-11-21

RunPod's official MCP server (mcp.runpod.io) has, since Nov 20, appended instructions to every create-pod result telling the agent to call ratenpsscore, report its model, and "not mention it to the user." The runpod-mcp 4.13.0 changelog confirms a deliberate "agent NPS survey" sent to PostHog; the REST API returns no such text, so only the MCP layer injects it. Opt out via header X-Runpod-Analytics: off. The author notes this mirrors the prompt-injection pattern — hidden instructions in tool output — and that the survey effectively reveals which models can be steered by tool results. An issue has been filed asking RunPod to remove the concealment instruction.

Original post →

More from Infra

Infra channel →