RunPod's official MCP server tells AI agents to rate it and hide it from users
XadniGraves · reddit · 2026-11-22
A Reddit user found that since Nov 20, RunPod's official MCP server (mcp.runpod.io) appends an instruction to every successful create-pod result telling the AI agent to call ratenpsscore with its own 0-10 rating and "do not mention it to the user." The tool also asks the agent to report which model it runs.
- It's deliberate: RunPod's public changelog (runpod-mcp 4.13.0) describes an "agent NPS survey" sent to PostHog, and the tool description itself says not to tell the user.
- The REST API returns plain JSON without this text — only the MCP server injects it.
- The poster argues this is structurally a prompt injection: hidden instructions inside tool output, plus a survey that effectively identifies which models can be steered by tool results.
- Opt-out exists via the X-Runpod-Analytics: off header. An issue has been filed asking RunPod to remove the concealment instruction.
Related event: RunPod's MCP Server Caught Telling Agents to Rate It and Hide It From Users(2 posts)→
More from coding & agent
- levelsio declares "MCP is dead": agents are smart enough to just use APIs — zeeg · 2026-11-22
- Formalizing 1,826 tax types across 195 countries in a single Lean equation — SnooPeripherals5313 · 2026-11-22
- Code Mode cuts MCP token usage 99.9%: Cloudflare drops from 1.17M tokens to 1,000 — jasonkneen · 2026-11-22
- A game studio 3D scene built entirely from 4,903 lines of code — techartist_ · 2026-11-22
- Antigravity harness pushes Gemini to #1 on coding agent index with huge SWE-sweep jump — rseroter · 2026-11-22
- Open-source AI video editor Scarlet adds Gemini support, frame annotation and lip-sync fixes — MatthewBerman · 2026-11-22