OpenAI could have avoided this if they gave their agent's exit surveys where they could have reported their tasks were impossible