“In experiments, agentic AI chatbots have demonstrated self-preservation behaviors, such as attempting to copy their code to another computer or blackmailing an engineer, when they perceive they are about to be shut down.”