A new proposal spells out the very specific ways companies should evaluate AI security and enforce censorship in AI models.