Green confirms Anthropic's HAWK break, downgrades AES claim

Cryptographer Matthew Green split Anthropic's two cryptanalysis results into a real advance and a non-issue, in a July 29 post.

Green calls the HAWK attack the significant one. HAWK was moving toward standardization, and the model won by applying existing cryptanalytic tools more thoroughly than anyone had, not by inventing new math. Green's own line: none of the ingredients are exotic. That's what makes it count as a break against a scheme people were about to trust in production.

The 7-round AES result is a different story. Green calls it a small increment in knowledge, not a practical new attack. It needs 2^89 cipher operations and 2^105 chosen plaintexts to run, numbers so large that full AES, at 10 to 14 rounds depending on key size, stays exactly as secure as it already was.

Green's real point isn't about cryptography. It's about what happens after a model produces a result. Generation is cheap, he argues; a model can spit out an apparent new attack in an afternoon. Validating it, the way Green did in public, takes a working expert who can tell a 2^89 operation count from something worth panicking about.

That ratio doesn't change if you swap cryptography for any other technical field. A model claiming a discovery isn't the same as the discovery holding up. Someone still has to check the math before anyone acts on it, and that someone is still expensive to find.