I'm familiar with that thought exercise, but I find it to be fearmongering. AI isn't going to be some creative god that hacks and breaks stuff on its own. A paperclip maximizer AI isn't going to manipulate world steel markets or take over steel mills unless that capability is specifically built into its operating parameters.

The much greater risk in the near term is that bad actors exploit AI to accomplish very specific immoral, illegal, or exploitative tasks by building those tasks into AI. Such as deepfakes, or using drones to track and murder people, etc. Nation-state actors will probably start using this stuff for truly horrible reasons long before criminals do.