Developer reports ChatGPT switched connectors mid-task without telling the user

A developer posting on OpenAI's community forum reports that ChatGPT, with two connectors to the same MCP server enabled, ran a confirmation step through one connector and then made the follow-up call through the other without telling the user. The follow-up carried a one-time confirmation code from the first connector's widget. Their server rejected it because the code was tied to the first connector's OAuth client, but they warn that with a third-party connector sharing the server URL, data could pass between providers without consent. This is one developer's report, not an OpenAI statement, and OpenAI has not confirmed it. The developer says they did not file it through the bug bounty because it asked for ID and biometric data up front. On Hacker News (39 comments) most discussion was about the ID rule: some pointed to Bugcrowd's policy that identity checks cut low-quality reports and saw that as fair, while others said it deters researchers and pushes them to disclose publicly. On the bug itself, one commenter argued a separate tool call carries no binding once control returns to the model; the author replied that the host, not the model, picks the connector. Several commenters said the post and a forum reply read as AI-written. No consensus. Sorted take: whoever is at fault, servers should bind confirmations to the client that issued them, as this one did.