“OpenAI's published model spec contains an "escape clause" stating that certain top-level priority policies, which overrule everything else, are not published and that the model is instructed to keep these policies secret from the user.”