AI cracks a post-quantum signature scheme, nudges 7-round AES — no new math required

Anthropic used Claude Mythos, its unreleased advanced model, to produce two cryptanalysis results. The stronger one is a key-recovery attack on HAWK, a proposed post-quantum signature scheme built on the module Lattice Isomorphism Problem. HAWK isn’t deployed or standardized, and the attack doesn’t transfer to the related Falcon scheme, but HAWK was well along in evaluation for a future standard. The attack stays exponential-time yet roughly halves the scheme’s security bits — fixable by doubling key sizes, except that HAWK exists only to be more efficient than the alternatives, so the patch guts its reason to exist. Crucially, the model invented no new mathematics; it just applied known tools more thoroughly than anyone had bothered to, which is precisely the kind of grind AI is good at. The result likely ends HAWK’s shot at standardization.

The second result, a faster attack on 7-round AES, sounds scarier than it is. Full AES runs 10 to 14 rounds and remains unbroken; 7-round attacks have existed since at least 2013, and this is a modest constant-factor improvement over that prior work. The numbers stay firmly theoretical — roughly 2^89 cipher operations after coaxing 2^105 chosen-plaintext encryptions from a victim — and because the attack can’t actually be run, the claimed speedup is a paper analysis that may or may not hold up once every detail is worked out. It’s legitimate technique-level progress, not a practical threat.

The process is the real story: Anthropic didn’t assemble a team of tuned domain experts, it essentially told the model to go find results and kept it working until it did, with little human intervention. That reframes the bottleneck as verification. Models are now good at producing real results but even better at producing results that merely look real, so human scrutiny matters more, not less. Full attacks like HAWK verify trivially — ship the code and watch it recover keys — while subtle speedup claims lean on Lean proofs or expert review, both sensitive to how the theorem is even stated.