Third-party dispatches, not graded receipts

Anthropic's Claude Mythos finds novel cryptographic weaknesses in HAWK and reduced AES

anthropic.com · 2026-07-30

Anthropic researchers used Claude Mythos Preview for ~60 hours of automated cryptanalysis, discovering genuine mathematical flaws in a post-quantum candidate algorithm and a weakened AES variant. The results, published in partnership with ETH Zurich and other universities, mark a notable demonstration of LLM-driven mathematical research, though neither finding impacts real-world systems today.

“Mythos Preview worked for 60 hours in total (~$100,000 in estimated API cost) and the main human interventions were to encourage it not to give up and "find something that worth publishing".”

source↗ · archived copy↗

← all dispatches