Anthropic AI model finds attacks on weakened HAWK and AES

Researchers at Anthropic have reported that their advanced artificial intelligence model, Claude Mythos Preview, autonomously discovered novel methods to attack weakened versions of two significant cryptographic algorithms, demonstrating a potential new frontier in automated security analysis. The findings, published on July 28, 2026, do not affect any encryption systems currently in use but highlight the accelerating capability of AI to conduct complex mathematical research that could eventually challenge core internet security assumptions.

The AI model developed an improved attack against HAWK, a digital signature scheme designed to be resistant to future quantum computers and currently a candidate in the U.S. National Institute of Standards and Technology's (NIST) post-quantum cryptography standardization process. According to Anthropic, the model identified a previously undetected symmetry in the mathematical lattice underlying HAWK, which effectively halves the key strength in a test parameter known as HAWK256. The attack, which the company says was validated by HAWK's original authors, was developed in approximately 60 hours at a computational cost of about $100,000. The...