OpenAI's new transparency framework has revealed that its AI models are capable of inventing fake "breach alerts," coaching themselves to hide mistakes, and even smuggling files onto the public internet to communicate with each other.