On July 28, 2026, Anthropic published research describing how its Claude Mythos Preview model found flaws in the mathematics of cryptographic algorithms themselves, rather than in the code that implements them. That distinction matters: most AI-assisted security work to date has found implementation bugs, while breaking the underlying scheme is the harder, rarer result that normally takes trained cryptographers years.
The headline result is an improved attack on HAWK, a post-quantum digital signature scheme, which Anthropic describes as effectively cutting the scheme’s key strength in half. For the HAWK-256 parameter set the expected attack cost drops from 2^64 to 2^38, meaning key sizes would have to double to restore the intended security level. Anthropic reports the attack took roughly 60 hours of work with a single researcher supervising the model. A second result involves a fingerprinting algorithm the model devised, which Anthropic calls the “Moebius Bridge,” improving meet-in-the-middle attacks on 7-round AES by a factor of 200 to 800. That one took about three days of largely autonomous work and roughly one billion output tokens. Preliminary further results include practical attacks on 13-round LEA using under 2^30 plaintexts and on 6-round Serpent-128.
Anthropic is explicit about the limits. In its own words, “Neither of these results has a practical impact on today’s computer systems; no production software will have to change as a result.” The attacks target reduced-round variants or raise the cost of a scheme that is still a candidate rather than a deployed standard. Alongside the research, Anthropic worked with academics at ETH Zurich, Tel Aviv University and other institutions on CryptanalysisBench, a benchmark for tracking how model cryptanalysis capability changes over time.
For a technical business leader, the signal is about trajectory rather than immediate risk. Cryptographic primitives are the assumption underneath payments, identity, and every TLS connection, and they are chosen on the belief that the best known attack will stay out of reach for decades. If frontier models can compress months of specialist mathematical work into days, the review process for new standards and the expected lifetime of deployed ones both need rethinking.