Claude Mythos Preview Independantly Cracks HAWK and AES Encryption

Anthropic's Claude Mythos Preview model found a way to weaken a post-quantum signature algorithm called HAWK, and also found attacks on a cut-down version of AES (a very common encryption standard — this wasn't the full AES people actually use). For HAWK-256, the model cut the attack's difficulty from 2^64 operations down to 2^38 — that's roughly 67 million times easier, using about $100,000 and 60 hours of compute. For the weakened AES version, a method called "Möbius Bridge" makes breaking it 200–800 times faster, again for about $100,000 and three days of compute.

None of this threatens anything you or I actually use — no software needs patching because of it. What matters to me is the shift happening underneath: an AI model pushing the cutting edge of cryptography on its own, in a field where humans used to need months to make this kind of progress. That's the part worth watching — the line between "AI as a tool you use" and "AI as a researcher doing its own work" is getting blurry.