Anthropic's Claude Mythos Preview model has achieved a breakthrough in cryptanalysis, discovering two new attacks: one on the post-quantum digital signature scheme HAWK and another on a reduced-round version of the AES-128 cipher. This is not just another test — it is a demonstration that modern AI agents are capable of conducting expert-level cryptographic research, previously accessible only to a narrow circle of specialists.
It is important to emphasize immediately: the vulnerabilities found do not threaten production systems and do not require urgent software updates. However, they show how far AI has advanced in understanding the mathematical structures underlying data protection.
Attack on HAWK: 2^26-fold reduction in security
HAWK is one of the candidates for standardizing post-quantum digital signatures, designed to verify transaction authorship without revealing the private key. Claude Mythos discovered a previously unknown symmetry in the mathematical structure of HAWK, which accelerated the key recovery attack.
According to my analysis, the effective security of HAWK-256 keys dropped from the expected 2^64 to 2^38 operations. This means that to maintain the previous level of protection, key sizes would need to be roughly doubled, negating one of the scheme's main advantages — its compactness. For HAWK-512 and HAWK-1024, the attack remains practically unattainable, confirming the localized nature of the vulnerability rather than a general collapse of lattice-based post-quantum cryptography.
Möbius Bridge: accelerating the attack on 7-round AES
The second result concerns the 7-round version of AES-128 — one of the most common symmetric ciphers, which uses a single secret key for encryption and decryption. The full version of AES-128 consists of 10 rounds, while the reduced version is used only in academic security tests.
The method, named Möbius Bridge, improved the known meet-in-the-middle attack on 7-round AES, providing a speedup of 200–800 times depending on the method of counting execution time. However, the scenario remains impractical: it requires 2^105 pre-selected plaintexts for encryption, making it a purely research result.
Results and prospects
Each of the two studies cost approximately $100,000 in API expenses. The search for the attack on HAWK took about 60 hours, while confirming the AES result required several hundred hours of researcher work. The company notified the algorithm authors, U.S. government agencies, and industry partners in advance.
My expert conclusion: This is not just a demonstration of AI capabilities, but a signal to the cryptographic community. If models at the level of Claude Mythos can find previously unknown symmetries in hours, then future standards must consider not only mathematical security but also resistance to automated analysis. HAWK may require revision, and AES-128 in its full version is currently safe, but the trend is clear: AI is becoming a full-fledged participant in cryptanalytic research.