Anthropic has announced that its advanced AI model, Claude Mythos Preview, has uncovered novel mathematical weaknesses in certain cryptographic algorithms, marking a notable step forward in AI-assisted crypto-analysis.
The findings, detailed in a recent research post from the company’s Frontier Red Team, highlight two primary advances that improve upon existing attacks but do not compromise any systems currently in widespread use.
The first discovery targets HAWK, a digital signature scheme designed to resist attacks from both classical and quantum computers.
HAWK is among the candidates advancing in the National Institute of Standards and Technology’s (NIST) post-quantum cryptography standardization process.
Despite undergoing extensive expert human scrutiny across two review rounds spanning roughly two years, the AI identified a previously unexploited structural property—a nontrivial symmetry within the underlying lattice mathematics.
This insight enabled a significantly more efficient key-recovery method. For the smallest parameter set, known as HAWK-256, the computational effort required to recover a secret key dropped dramatically from an estimated 2^64 operations to approximately 2^38.
The practical effect is a substantial reduction in the scheme’s effective security margin, roughly halving its key strength.
Addressing this would require larger keys, which could diminish some of the performance advantages that made HAWK appealing as a candidate. Importantly, HAWK has not been deployed in any production environments, so the finding poses no immediate real-world risk.
The discovery process involved roughly 60 hours of semi-autonomous work guided by a single researcher.
The second advance concerns a reduced-round version of the Advanced Encryption Standard (AES), the globally dominant symmetric cipher standardized by NIST in 2001.
Cryptographers routinely study simplified variants with fewer rounds than the full ten-round AES-128 to probe potential weaknesses.
Claude Mythos Preview developed a new technique, referred to as the Möbius Bridge, that eliminates one of the guessing steps in prior meet-in-the-middle approaches.
Combined with other optimizations generated by the model, this yields an attack that is between 200 and 800 times faster than the previous best results on seven-round AES, depending on measurement criteria.
The full-strength AES remains unaffected, and the theoretical nature of the improved attack—requiring an impractically large volume of data—means it has no bearing on operational systems.
Both results emerged largely through autonomous exploration by the model, supported by computational tools, literature access, and limited human scaffolding rather than deep specialist direction in every step.
Each effort incurred substantial computational costs on the order of $100,000 in API usage, with additional extensive human verification required afterward, particularly for the AES work.
Anthropic has also noted preliminary results involving reduced-round versions of other algorithms, such as the Lightweight Encryption Algorithm and Serpent, though these remain under further investigation before full publication.
These developments underscore the growing capacity of frontier AI systems to contribute meaningfully to cryptography research by identifying subtle mathematical vulnerabilities that can evade prolonged human analysis.
Rather than signaling imminent threats to deployed encryption, the work is framed as a constructive tool for strengthening future standards and accelerating defensive research. Anthropic plans additional disclosure of complete technical details following further review.