Claude Just Did 2 Years of Crypto Review Homework in Days
What happened
Anthropic says its Claude Mythos Preview model discovered two new cryptographic weaknesses: it significantly weakened HAWK, a post-quantum digital signature scheme under NIST review, and found a faster attack on round-reduced AES. The HAWK attack took Claude just 60 hours and effectively halved the scheme's key strength — after the algorithm had already survived two years of expert human scrutiny.
Why this matters
Cryptography has long been considered one of the last bastions where human mathematical intuition reigns supreme. If AI models can meaningfully speed up attacks on candidate algorithms before they're even deployed, that's either a huge win for defensive security research or a preview of a much scarier offensive capability curve.
The slightly cynical read
Anthropic conveniently frames this as "cryptography research working as intended" rather than "our model can now break things humans spent years failing to break." Both are true, but only one makes for a comforting blog post headline.
What to watch next
Neither HAWK nor the AES variant is in production, so nothing breaks today — but expect NIST and other crypto bodies to start stress-testing candidate algorithms against frontier models before standardization, not after.