Shows the superficiality of training in censorship / alignment. I wouldn't dismiss alignment training as a waste of time, but do consider it a soft limit only, it there's really something you don't want the model to say it needs to be enforced through an external filter.