San Francisco, Anthropic researchers have used the Claude Mythos Preview artificial intelligence model to identify improved attacks against two cryptographic algorithms, demonstrating the emerging potential of AI-assisted security research.
The first finding relates to HAWK, a post-quantum digital-signature scheme being evaluated by the US National Institute of Standards and Technology. Despite undergoing two years of expert review, Claude Mythos developed an improved key-recovery attack in approximately 60 hours.
The attack identifies a previously unexploited mathematical symmetry in the lattice used by HAWK. Anthropic said the finding effectively reduces the scheme’s key strength by half, meaning its key sizes would need to be doubled to provide the originally expected level of security.
For the smaller HAWK-256 configuration, the estimated computational cost of a complete key-recovery attack was reduced from 2642^{64}264 to 2382^{38}238. The finding is specific to HAWK and does not affect other post-quantum signature candidates or lattice-based cryptography generally.
The second finding concerns a reduced-round version of the Advanced Encryption Standard. AES-128 normally applies 10 encryption rounds, while the research examined a deliberately weakened seven-round version used for academic cryptanalysis.
Claude developed a new fingerprinting technique, called the Möbius Bridge, which improved the speed of the previously strongest attack by between 200 and 800 times. However, the method does not break the full 10-round AES cipher used in production systems.
Anthropic said neither discovery currently affects deployed software or online services. HAWK remains a candidate scheme and has not been deployed, while the AES finding applies only to a simplified research version of the cipher.
Claude conducted much of the research autonomously, although human researchers subsequently spent substantial time validating its conclusions. Anthropic estimated that developing each of the two primary attacks involved approximately $100,000 in model usage costs.
The company followed coordinated disclosure procedures and shared the HAWK finding with its authors before publication. It also consulted academic, government and industry experts to validate the results.
Anthropic is an artificial intelligence research and product company that develops the Claude family of AI models. Its Frontier Red Team studies advanced model capabilities and risks across cybersecurity, autonomous systems and other dual-use domains.
