Claude found new cryptographic attacks, with carefully bounded consequences

Anthropic reports that Mythos Preview found a mathematical weakness in HAWK, a candidate post-quantum signature scheme, after about 60 hours of work. The attack reduces the scheme’s effective security and undermines some of the advantages that made it attractive.

A separate result improves attacks on a version of AES with fewer rounds than the full cipher, by a reported factor of 200–800. Researchers deliberately study such weakened variants to understand the security margin of the complete design.

HAWK is still a candidate, and the AES result does not break the full algorithm. Anthropic says neither finding requires changes to production systems.

Simon’s attention goes to the human interventions. The researchers repeatedly pushed the model to pursue a genuinely difficult result instead of settling for an easier target or deciding the task was impossible. The published prompts are informal and sometimes misspelled; what matters is the research objective and the insistence on checking a worthwhile claim.

That offers a concrete view of AI-assisted cryptanalysis: literature review, mathematical exploration, computation, and validation, supported by human direction. The HAWK work’s estimated API-equivalent cost was around $100,000. The accompanying papers and code make the result inspectable beyond an announcement.