A post titled LLMs won't break symmetric crypto generated substantial discussion this week — 97 comments on 75 points, a ratio that usually means disagreement.

The argument is worth understanding not for the conclusion, which is not controversial among cryptographers, but for the structure of the reasoning, which is unusually clean.

Why symmetric crypto is a special case

Most capability questions about language models are empirical and open. Can they do research-level mathematics? Maybe, partially, depends on the field. Can they write correct systems code? Sometimes, with review.

Symmetric cryptography is different, because the security argument does not depend on the attacker's cleverness. A well-designed block cipher is built so that recovering the key requires searching a space of size 2^128 or larger, and that bound holds against any attacker, including one with excellent pattern recognition.

The relevant question is not "is the model smart" but "does the model change the size of the search space." Language models do not. They are not a new computational model; they run on the same hardware as everything else.

Where the intuition goes wrong

The reason the question recurs is that language models are extremely good at a class of problem that looks similar: finding structure in sequences that appear random to humans.

That skill is genuinely relevant to broken or homemade ciphers, which leak statistical structure. It is irrelevant to AES, whose entire design goal is that no such structure exists to find.

Confusing the two is easy because both are described as "code-breaking."

Where AI actually threatens cryptographic systems

The realistic risks are not in the primitives:

• Implementation flaws. Side channels, weak randomness, key reuse, padding oracles. Models are already useful at finding these, because they are software bugs and models are good at software.
• Protocol misuse. Correct primitives assembled incorrectly. Extremely common and very much a pattern-matching problem.
• The human layer. Phishing and social engineering at scale, which is where the actual key material usually leaks from.
• Harvest-now-decrypt-later, which is a quantum concern rather than an AI one, and applies to asymmetric rather than symmetric cryptography.

Every item on that list is a real security problem. None involves breaking the cipher.

The generalisable lesson

The interesting property of this argument is that it distinguishes capability limits imposed by mathematics from limits imposed by current technique.

Most claims about what models cannot do are of the second kind, and they age badly — the field has retired a long list of them. Symmetric cryptanalysis is of the first kind. The bound comes from the size of the key space, not from the state of the art, and no amount of scaling moves it.

Being able to tell those two categories apart is one of the more useful skills for reading capability claims, and there are fewer clean examples of the first category than the discourse suggests.