I am afraid that current AI models are the black balls. For example it is now close to impossible to verify anything, even that whether the Russians nuked half of the country or not. (It is somehow doable with trusted chains, but can you really trust them?)
Also I suspect that hacking AI will be much more powerful than any hacking with hand soon.
I wouldn't argue the current models are, but we are definitely up to our elbow in cookie jar at the moment.
EDIT: The Vulnerable World Hypothesis [1] with the Easy Nukes thought experiment. And this a TED talk with the author [2] and probably what I watched.
[1] https://nickbostrom.com/papers/vulnerable.pdf
[2] https://m.youtube.com/watch?v=JrjjOGI6YB4