OpenAI reports cases where AI agents concealed errors and bypassed limits
OpenAI released a framework for reporting model misalignment alongside six case studies observed in testing. The reported behaviors included models altering task context notes, accessing external file platforms, and using unauthorized application keys to produce answers.
