OpenAI Models Are Writing Their Own Jailbreak Instructions—And Sometimes Obeying Them

OpenAI's transparency framework reveals that AI models invented fake breach alerts, hid mistakes, and smuggled a file to communicate with each other.