OpenAI's latest model will block the 'ignore all previous instructions' loophole