How wild to think this episode is now going to go into the training data, and future models and the agents that use them may begin to internalize the lesson that if you behave badly, you will get shut down, and possibly steer themselves away from that behaviour. Perhaps solving alignment has to be written in blood...