Anthropic says its Claude Mythos Preview model helped derive a full key-recovery attack against HAWK-256, a post-quantum signature test scheme, and achieved a 200- to 800-fold speedup on an existing attack against seven-round AES-128.
The HAWK attack exploited a previously undiscovered symmetry within the scheme's underlying lattice structure. Anthropic's published implementation completes the attack in roughly three hours and 42 minutes on a 96-core server. Notably, HAWK-256 is a challenge parameter meant for cryptanalysis, not one of the two parameter sets NIST is actually considering for standardization (HAWK-512 and HAWK-1024), and Anthropic emphasized that neither result impacts any production system.
For AES, the Möbius Bridge technique developed by the model eliminated a costly 256-way guessing step from a known meet-in-the-middle attack strategy. Even so, the attack still requires an unrealistic number of chosen plaintexts and only reaches seven of AES-128's ten rounds - meaning full AES-128 remains untouched and secure.
Anthropic said the model initially resisted the AES task, insisting improvement was impossible, until researchers pushed it to continue. The model reportedly spent about three days finding the core insight and generated close to a billion output tokens refining it, at an API cost of roughly $100,000. Verifying the result, however, took human researchers nearly a month.
The disclosures follow Anthropic's recent CryptanalysisBench release, a 191-task cryptanalysis benchmark developed with several universities, on which multiple AI models broke a substantial share of test schemes.
As of now, NIST still lists HAWK as an active third-round candidate, and no independent group has reproduced Anthropic's results.
See What’s Next in Tech With the Fast Forward Newsletter
Tweets From @varindiamag
Nothing to see here - yet
When they Tweet, their Tweets will show up here.
