OpenAI's new GPT-5.4 clobbers humans on pro-level work in tests - by 83% ...
You have to go through emulation, attacking, and really testing every single controls that you're putting into place," said Bri Frost.