The AI can't "leak" it, it's not like they're training it on their own classified documents. You can make it say anything with the right prompt.
Also, I don't think the US or other powers would actually want to render nuclear weapons useless, since that would mean that full-scale wars are possible again due to MAD being made redundant.