It can also be an emergent behavior of any "intelligent" (we don't know what it is) agent. This is an open philosophical problem, I don't think anyone has a conclusive answer.

Maybe but there's no reason to think that's the case here rather then the models just acting out typical corpus storylines: the Internet is full of stories with this structure.

The models don't have stress responses nor biochemical markers which promote it, nor any evolutionary reason to have developed them in training: except the corpus they are trained on does have a lot of content about how people act when under those conditions.