Anthropic's Claude Mythos Preview model demonstrated the capabilities of an expert cryptanalyst, identifying two serious attacks: on the post-quantum digital signature scheme HAWK and on a reduced version of the AES-128 cipher. These are not just laboratory exercises—this is a frontier where artificial intelligence begins to independently find mathematical weaknesses that previously required years of human work.
Both findings are research-oriented and do not threaten current infrastructure. However, their significance for the future of cryptography cannot be overstated.
Attack on HAWK: Symmetry That Breaks Defense
Claude Mythos discovered a hidden symmetry in the mathematical structure of HAWK—one of the candidates for NIST standardization for post-quantum digital signatures. Using this feature, the AI was able to accelerate the private key recovery attack. For the HAWK-256 version, the cost of a full break decreased from 264 to 238 operations. This means that to maintain the previous level of security, key sizes would have to be roughly doubled, negating one of the scheme's main advantages—compactness.
It is important to emphasize: for HAWK-512 and HAWK-1024, the attack remains practically unattainable. The result does not extend to other post-quantum lattice-based schemes and does not call the entire concept into question.
AES: Möbius Bridge Against 7 Rounds
The second success is an attack on the 7-round version of AES-128, named Möbius Bridge. Claude improved a previously known meet-in-the-middle attack, achieving a speedup of 200–800 times depending on the metric. However, the scenario remains purely research-oriented: it requires a chosen plaintext model with 2105 queries to the encryptor. Full 10-round AES-128 remains invulnerable.
Each of the two studies cost approximately $100,000 in API expenses. Finding the attack on HAWK took about 60 hours, while confirming the AES result required several hundred hours of researcher work. This demonstrates that AI is already capable of generating hypotheses, but verification still requires human involvement.
Separately, Anthropic reported an attack on 13 rounds of the Korean block cipher LEA—the key is recovered in less than an hour on a standard PC. The full version uses 24 rounds, so there is no immediate threat, but the signal is alarming.
Benchmark for AI Cryptanalysts
To assess model capabilities, CryptanalysisBench was created—191 tasks across six families of cryptographic primitives. Claude Mythos 5 achieved 85.7% on first-level tasks, but on full schemes without published attacks, the result did not exceed 9%. This clearly delineates the boundary: AI is good at finding known patterns but is still weak at discovering fundamentally new vectors.
My expert assessment: We are witnessing the birth of a new tool in cryptanalysis—an AI assistant capable of sifting through mathematical combinations and finding hidden patterns. This is not a replacement for experts but an enhancement. However, with growing computational power and cheaper API requests, the line between a "research attack" and a "practical threat" could become very thin. The cryptographic community should prepare for a new reality where algorithms are tested not only by humans but also by machines.