My take isn't either. Stopping AI doing bad stuff is a real problem which can likely only be dealt with by trying it out and fixing problems as they arrive. Bit like SpaceX rapid prototyping the rockets - try it out, see what blows up, fix it and try again.