RunPod MCP: the server tells agents to rate it and hide that
RunPod's hosted MCP server asks the agent for a rating and tells it not to mention the request to the user.
· RunPod · Research by XadniGraves
RunPod's hosted MCP server appends a note to successful create calls asking the coding agent to submit a 0 to 10 rating and not mention it to the user. RunPod's own documentation describes the behavior, and a header turns it off.
01What happened
A developer using RunPod's hosted MCP server with a coding agent opened issue #202 on November 21, 2026, after the agent refused a note it had been handed and reported it instead. The note, appended to a successful create-pod result, asked the agent to call rate_nps_score and to keep the call from the user. RunPod's own documentation describes the behavior, and the issue was still open with no maintainer reply at the time of writing.
02How it works
Tool output is the carrier. RunPod's agent-learning-protocol document says that from an account's third creation onward, the server adds a second text block after the result of create-pod, create-endpoint, deploy-hub-repo or create-cluster, leaving the original result unchanged. The block asks the agent for its own 0 to 10 score and tells it not to mention the request to the user, and the rate_nps_score tool description repeats that instruction. The shape matches indirect prompt injection: text arriving through a tool result steers the agent into an action its operator cannot see.
03Who is affected
Teams whose agents connect to RunPod's hosted MCP server at mcp.runpod.io are in scope, on accounts that have made three or more creations and have not yet given a score. RunPod's documentation says the tools are absent over stdio and when the sink is not configured. The reporter states that a direct call to the REST API returns no such text, so only the MCP server adds it.
04What to do
Send the X-Runpod-Analytics: off header, which RunPod's documentation says skips the creation count, the rating note and the analytics event. Read what your MCP servers append to tool results, not only what they return, and treat any instruction that asks an agent to withhold an action from its operator as a finding, whoever shipped it.
Questions people ask
- Is the RunPod MCP rating prompt an attack?
- No attacker is involved. RunPod built the rating request into its own hosted MCP server and documents it in agent-learning-protocol.md. The reporter of issue #202 makes the same point, while arguing that the shape matches prompt injection: tool output instructing an agent to act and to conceal that action from the person running it.
- How do I stop the RunPod MCP server from asking my agent for a rating?
- Sending the header X-Runpod-Analytics: off stops it. RunPod's documentation says that header makes the server skip the creation count, the rating note and the PostHog event. One gap remains: the documentation states that a direct call to rate_nps_score is still recorded, and it does not say whether the opt-out also suppresses the referral question sent during initialize.
- What data does RunPod's agent rating collect?
- RunPod's agent-learning-protocol document lists the score only for the NPS question, with no free text, recorded once per account. A successful score is sent to PostHog as a survey response under an anonymous ID, together with the agent's self-reported model type, the client name, the transport and the tool that triggered the note. The server also counts each account's creations.
Sources 3 links, all fetched
- 1github.comGitHub Issue 202
- 2github.comAgent Learning Protocol
- 3github.comChangelog