Anthropic’s AI, Claude Mythos, managed to reveal fresh vulnerabilities in two cryptographic algorithms that have been under scrutiny for years. While these discoveries don’t impact current encryption systems, they highlight how advanced AI can uncover mathematical weaknesses that human experts previously missed.
The AI found a major weakness in HAWK, a post-quantum digital signature scheme still being tested by the National Institute of Standards and Technology. Claude identified a hidden symmetry in HAWK that allowed for a much faster key recovery attack, effectively halving the algorithm’s security for its smallest configuration. For HAWK 256, the computational effort to break the key dropped from 2^64 to 2^38 operations, which is a massive reduction in difficulty. Although this attack is not practical against larger keys and doesn’t affect other post-quantum candidates, it means HAWK will need to double its key sizes to keep the intended security level, sacrificing some efficiency.
What’s impressive is that Claude developed this attack largely on its own in about 60 hours, using a combination of mathematical software, research papers, and computation tools. A human researcher only gave occasional hints and didn’t have deep expertise in lattice-based cryptography. The project cost around $100,000 in API usage.
On a separate front, Claude improved an attack against a reduced version of the Advanced Encryption Standard (AES), specifically a seven-round AES 128 cipher used for research but not in the wild. By introducing a new fingerprinting method called the Möbius Bridge, Claude eliminated a guess step in the attack, speeding up the process by a factor between 200 and 800 compared to previous techniques. This breakthrough came from about a billion tokens generated by Claude while refining the approach.
These findings illustrate how AI is rapidly becoming a powerful tool in cryptanalysis, pushing the limits of what’s possible in algorithm testing. It also raises questions about the future design of secure systems as AI continues to advance.
This material is for informational purposes and does not constitute financial advice.
