Discovering Cryptographic Vulnerabilities Using Claude

Anthropic's Claude Mythos Preview achieved results that go beyond finding conventional software implementation bugs, discovering mathematical flaws that exist within the cryptographic algorithms themselves.

The first major achievement was improving attack performance against the HAWK digital signature scheme, a post-quantum cryptography (PQC) candidate. Mythos found an attack method that halves HAWK's key strength in just 60 hours.

The second achievement was an attack on a variant model of AES, the most widely used symmetric cipher. Mythos discovered a vulnerability in round-reduced AES, improving speed by 200 to 800 times over the previous best-known attack.

These discoveries do not directly impact currently operating production systems. HAWK is still at the candidate stage, and the AES attack targeted a reduced version rather than the full cipher.

This research demonstrates that frontier AI models have strong potential as stress-testing tools for verifying the stability of cryptographic algorithms and enhancing security both before and after real-world deployment.