who? the state on advice of the intelligence community: machine learning inference bruteforcing not bits in private keys, but bruteforcing proof search space modulo machine intuition.

At some point cryptography evolves faster than human cryptographers can follow, at that point anyone can only trust cryptographic schemes proposed by their own LLM models, since they will be direct witness to useful responses for apparently unbreakable cryptoschemes with a mathematically known backdoor for the designing LLM.

Why would one power block accept usage of LLM-derived crypto from another power block, if they know it is likely backdoored? The same within a power block: why would one ally financially trust crypto from another ally in the same power block, if they are witness to how easy it will be to request new convincing crypto schemes with a backdoor? The same within a single nation state: why would one bank accept the cryptography proposed by another bank, if they are witness to how easy it would be to ensnare one another with LLM-designed crypto schemes with backdoors.

Its not that humans are incapable of pulling power plugs when the machines start organizing human-on-human violence, it's that no side, no party feels like they can afford a gentleman's agreement to not consult LLM's to find vulnerable hardness assumptions in mathematical cryptography. Perhaps many powerful entities would be more than willing to pull such a plug, but not before others have provably demonstrated pulling the plug,... which is basically waiting until no other human factions exist.

And there's who pays it, and they will pay private companies to compete in order to assure competitive prices for the government.