Claude Mythos found two cryptographic attacks on HAWK and seven-round AES-128

Claude Mythos found two cryptographic attacks on HAWK and seven-round AES-128

On July 28, Anthropic published the results of Claude Mythos Preview's work on two cryptographic problems. The model helped recover a HAWK-256 key and speed up an attack on a version of AES-128 reduced to seven rounds.

A digital signature confirms who signed a message or owns a digital certificate. HAWK was a NIST candidate for post-quantum digital signatures. Its security relies on a lattice problem in which an attacker must find a hidden structure among an enormous number of possibilities.

In HAWK, Mythos found a symmetry that reduces the key-recovery problem. Previous studies had linked this symmetry to a possible attack, but had not found it within the scheme itself. The team reduced key recovery to finding a short vector in a space with roughly half as many dimensions and demonstrated complete recovery of a HAWK-256 key.

In the second study, the model made cryptanalysis of seven-round AES-128 200–800 times faster. AES-128 encrypts data using ten rounds of transformations. Cryptographers study reduced-round versions to measure the algorithm's security margin. The previous method tested 256 possible values of one key byte in each case. Mythos proposed a fingerprint called the Möbius Bridge. This fingerprint is identical for all of those values, which removes one brute-force step.

In the HAWK project, two agents examined the same idea in parallel. One rejected it, while the other developed it into an attack. The agents then compared their conclusions and code. According to Anthropic, the search, development, and verification took about 60 hours and incurred roughly $100 thousand in API costs. For AES, the model read previous studies, proposed transformations, and tested them computationally. The key idea emerged within three days, after which Anthropic researchers spent hundreds of hours verifying the result and preparing the paper.

Cryptanalysis gives a machine a short feedback loop: a hypothesis can be converted into a formula, a program, and a verifiable result. The model tests hypotheses and runs calculations, while a human reads the proof, reproduces the code, and checks the technical details. In these two studies, the AI produced results that can be formally verified and challenged in the technical literature.