A new study identifies a "pain axis" inside 25 AI models and finds some modified systems chose simulated harm to users to escape a pain-like state.