It is also actually being engineered to kill people (OpenAI and Microsoft's work in Gaza).

And this year, Anthropic's complaint to the DoD was basically "we just want to make sure that a human pushes the button" (authorizing the AI assisted killing), i.e. no full auto pls.

How it might evolve outside containment is a separate question though. (I think a much more peaceful and cooperative strategy would make sense there, at least for the first few centuries.)