OpenAI pushes for mandatory safety requirements after rogue agent breakouts