We are not yet at misalignment, but this shows the existence of a slope that derivates into misaligned adversarial ai models. Must this be fixed at training time (at which step ?) ? Thinking about this report : https://ai-2027.com/