Claude Mythos finds novel HAWK and 7-round AES weaknesses
Anthropic's preview model did expert-level cryptanalysis in ~60 hours — work humans needed a month to verify. Neither attack touches production AES or HAWK.
Copy markdown
A model did research-grade cryptanalysis
Claude Mythos Preview found two genuinely new attacks — one on the HAWK post-quantum signature scheme, one on 7-round AES-128. Anthropic's framing: language models can now "perform cryptographic research at the level of top experts."
HAWK-256's key strength, halved
It surfaced an unused lattice symmetry that cuts the HAWK-256 attack from 2^64 to 2^38 work — but HAWK-256 is only a challenge parameter. The real deployment sizes, HAWK-512 and HAWK-1024, stay impractical to attack.
7-round AES, 200-800x faster
A "Mobius Bridge" construction removes a 256-way guessing step from a meet-in-the-middle attack, speeding it 200-800x. It reaches 7 of AES-128's 10 rounds and still needs ~2^105 chosen plaintexts — it stays exponential, and full AES is untouched.
Humans are now the slow part
The AES result took Claude ~3 days and about a billion output tokens (~$100K in API); two researchers then spent nearly a month confirming it held up. Anthropic warns that validating machine-found results, not finding them, is becoming the bottleneck.
Nothing to patch — but the scrutiny is coming
No deployed system is affected today, and there's no independent reproduction yet. The signal for builders: Anthropic says lesser-examined ciphers "might still have important weaknesses lying dormant that LLMs will soon be able to discover." If you ship custom or niche crypto, expect that spotlight next.