Anthropic has announced that one of its experimental AI systems, Claude Mythos Preview, has achieved two notable breakthroughs in cryptography research by identifying previously unknown mathematical weaknesses in encryption algorithms.

Unlike earlier AI-assisted security research, which mainly focused on finding software implementation bugs, this work targeted the cryptographic algorithms themselves—a much more challenging problem.

Breaking New Ground in Cryptanalysis

The first discovery involves HAWK, a post-quantum digital signature scheme currently being evaluated as part of the U.S. National Institute of Standards and Technology (NIST) process for future cryptographic standards.

Working with minimal human guidance, Claude developed a more efficient attack that significantly reduces the theoretical security margin of HAWK. Although the attack remains impractical for real-world exploitation, it suggests that the proposed key sizes may need to be reconsidered before the algorithm can be adopted as a standard.

The second result focused on a 7-round version of AES-128, a simplified variant of the Advanced Encryption Standard commonly studied by cryptographers. Claude discovered a new optimization technique that makes the best-known academic attack on this reduced version between 200 and 800 times faster than previous approaches.

Importantly, this does not compromise the full 10-round AES-128 algorithm used in modern applications. Today’s encrypted communications remain secure.

AI as a Research Partner

According to Anthropic, the AI completed much of the research process autonomously, including reviewing academic papers, proposing hypotheses, testing mathematical ideas, and validating its own findings.

Researchers estimated that developing each result required roughly a week of computation and significant cloud resources, while human experts spent considerably more time verifying the AI’s conclusions than the AI spent generating them.

Why It Matters

Although these discoveries have no immediate impact on production systems, they highlight a rapidly changing landscape for cybersecurity and cryptography.

For decades, discovering weaknesses in encryption algorithms has required years of work by specialized researchers. Frontier AI models are now demonstrating the ability to accelerate parts of that process, helping identify design flaws before they become widely deployed.

This capability could strengthen future cryptographic standards by allowing researchers to stress-test new algorithms more thoroughly during their development.

Looking Ahead

Anthropic has also introduced CryptanalysisBench, a benchmark designed to evaluate how well AI models perform cryptographic research. The initiative aims to help researchers measure future advances while encouraging responsible disclosure and collaboration with the academic community.

As AI continues to evolve, its role in cybersecurity is expanding beyond vulnerability detection. It is increasingly becoming a valuable research assistant capable of tackling complex mathematical problems that were once considered the exclusive domain of cryptography specialists.

While AI is unlikely to replace human cryptographers anytime soon, this research demonstrates that it may soon become one of their most powerful tools.