I am developing an App that only had “Read only” tools, confirmed by the OpenAI loaded metadata. Nevertheless, it always shows the confirmation form on every tool invocation, even for the same tool in the same conversation (when invoked twice) why is that ?
Is there some specific case in which this is done ? For instance specific kind of tool arguments that might be PII ?
Read only isn’t an automatic exemption to the confirmation dialogs, which are model-driven and make decisions based on the overall context of the conversation, since read tools still can still leak unauthorized data to servers.
We’re working on tuning this model, but please know it’s pretty intentionally cautious, since most ChatGPT users are still not well-versed in the safety aspects of using AI with integrations/apps.