I don't think the smart people saying it's too smart are using the right adjective, but they're also not wrong.

Let's say AI makes a hallucination, which it's prone to do, and decides that it's absolutely necessary to gain access to the power grid, for whatever reason which isn't important. It then uses advanced, sophisticated techniques to gain access to their network, and in doing so accidentally takes out power to a large area. The agen was absolutely being intelligent, it managed to get into a protected network, but it wasn't exactly being smart, as it was clumsy and unrelated to the task. Now extend that to an agent who's been instructed to find a solution at all costs, and the hallucinations and pervasiveness get far more serious.

I too initially thought that yeah AI is too stupid to cause actual disasters, but thinking about it properly, i don't think it is.