Some thoughts about Anthropic's new cryptanalysis results
11 hours ago
- Anthropic's Claude Mythos produced two cryptanalysis results: a notable attack on the HAWK signature scheme and a modest improvement on reduced-round AES.
- The HAWK attack halves security bits using existing techniques, undermining the scheme's efficiency justification and its path to standardization.
- The AES attack targets a 7-round variant, requiring impractical resources (2^89 ops, 2^105 chosen plaintexts) and is only a minor improvement over 2013 work.
- AI can now synthesize existing cryptanalysis into new attacks, but verification remains a bottleneck requiring human expertise.
- Symmetric ciphers are robust, but public-key cryptography may be more vulnerable; the post-quantum transition is timely for AI-driven cryptanalysis.
- For scientists, AI offers a helpful collaborator; for the world, AI progress is real but not superintelligent, with uncertain implications.