There's no AI safety. Either the AI does what the user asks and so the user can be prosecuted for the crime, or the AI does what IT wants and cannot be prosecuted for a crime. There's no safety, you just need to decide if you're on the side of alignment with humans or if you're on the side of the AIs.