Machine Learning Model Cracks Cryptographic Systems at Unprecedented Pace

Anthropic's Claude Mythos Preview model has shaken up the world of cryptography by demonstrating novel attacks on two prominent cryptographic systems, HAWK and AES. The company claimed that its model found a symmetry in the mathematical lattice used by HAWK, a post-quantum signature scheme submitted to NIST's competition for post-quantum cryptography.

The attack on HAWK effectively halved the scheme's security by reducing the key size from 256 bits to 128 bits. This can be fixed by doubling the key size, but that would negate most of the reasons why HAWK was chosen in the first place. As for AES-128, Mythos improved an attack on a reduced-round version of the cipher, making it significantly faster and more efficient.

The most striking aspect of this story is not the technical results themselves, but rather the fact that they were achieved at a cost of $100,000 each. This is a fraction of what a single postdoc researcher might earn in a year. The model's ability to produce novel attacks at an unprecedented pace has significant implications for the field of cryptography.

The article also highlights the growing concern about verification in cryptography. While discovery got cheaper with the advent of machine learning models like Mythos, validation and verification remain a bottleneck. Human researchers may struggle to keep up with the sheer volume of results produced by these machines, leading to potential security risks.