What was shown, and within which limits
Anthropic reports two results. The first improves the best known attack on the post-quantum signature scheme HAWK: the expected attack cost for HAWK-256 drops from 2^64 to 2^38, the outcome of 60 hours of work. The second is a new attack on seven of AES-128's ten rounds, 200 to 800 times faster than the previous best. Both were produced with Claude Mythos Preview.[1]
More decisive than the exponents are the scope conditions. HAWK is a third-round candidate in NIST's additional digital signatures competition, not a deployed scheme. The AES result applies to a round-reduced version and does not break the full cipher. Anthropic's own wording preserves those limits: no production system is affected, and the HAWK attack remains exponential-time rather than polynomial-time. Halving the key strength of a signature scheme is a serious cryptanalytic result; it is not the breaking of a deployed system.[1]
The distance between generation and validation
The most informative number in the write-up is not 2^38. The HAWK result came out of 60 hours of work; on the AES work the model produced roughly a billion tokens over three days; each attack cost about $100,000 in API credits. Against that, it took two researchers nearly a month to gain confidence that the discovered method was correct. Generation is measured in days, validation in expert weeks.[1]
That ratio makes the result a finding about method rather than a headline about capability. If candidate cryptanalysis can be produced faster than experts can check it, the binding constraint in this area moves from idea generation to verification capacity — and verification depends on expert time, the hardest input to scale. A narrower reading fits the same evidence: these are two unusually intricate results selected for publication, and a month of checking may describe their difficulty rather than any general ratio. What would separate the two readings is an independent re-derivation, and a report of how long it took.[1]