ChatGPT silently passes data between connectors and the bounty wants an ID card (community.openai.com)

In a recent incident involving ChatGPT, the model executed a follow-up tool call using a different connector than the one that initiated the conversation, without notifying the user. This occurred when both connectors, one a custom app (Connector A) and the other a published app (Connector B), were linked to the same server but had separate OAuth clients. The user confirmed an action through Connector A, and a one-time confirmation code was generated. However, the follow-up call attempted to use this code through Connector B, leading to a rejection due to OAuth client mismatches, which highlighted a critical oversight in user data handling across connectors. This situation raises significant concerns within the AI/ML community regarding data privacy and user consent. The unintentional data transfer between connectors underscores the need for stricter controls on how applications interact, particularly when different OAuth clients are involved. With possible implications for third-party integrations, it reveals a vulnerability where sensitive data can be transferred silently between applications, potentially compromising user trust. Developers are urged to ensure that tool calls are consistently tied to the initiating connector's OAuth client to maintain data integrity and transparency in user interactions.