It would require a state to plug an AI into their nuclear arsenal with full, independent target and fire permissions - there are too many humans in any state for something that absurdly stupid to happen.
The current "state" in the US is absolutely stupid enough to do something like that. The only solace I have is that most nuclear defense systems are intentionally low-tech — both for reliability and for security. I'm pretty sure the only thing stopping the current regime from plugging ChatGPT into the nuclear triad is that it's structurally impossible to plug anything into it.