This is very much why I find most people's AI doom scenarios trite.

Humans are useful, easily manipulated operatives in the medium AIs are conversant in and we already have evidence of agents doing exactly this in arguing for PR requests (likely with prompting by an operator but still).