Artificial intelligence continues to demonstrate capabilities far beyond simple text generation. My analytical team has recorded a significant breakthrough: the Claude Mythos Preview model successfully identified two cryptanalytic attacks — one on the post-quantum digital signature scheme HAWK and another on a truncated version of the symmetric cipher AES-128.

This is not about breaking working systems, but about fundamental research results. However, their significance cannot be overstated: AI has for the first time demonstrated the ability to conduct expert-level cryptanalysis, which was previously the exclusive domain of humans.

Attack on HAWK: Security reduced, but not broken

Claude Mythos Preview discovered a previously unknown symmetry in the mathematical structure of HAWK — a candidate for post-quantum digital signature standardization. This finding accelerated a key recovery attack. For the HAWK-256 version, the effective key security dropped from 2^64 to 2^38 operations. This is a serious weakening, but for HAWK-512 and HAWK-1024, the attack remains practically unfeasible.

In my assessment, this discovery does not spell the end for lattice-based post-quantum cryptography as a whole, but it forces a reconsideration of security parameters for schemes with small keys. Anthropic rightly notes that to maintain the previous level of protection, HAWK key sizes would need to be roughly doubled, which negates one of the scheme's key advantages — compactness.

Möbius Bridge for AES: Hundreds of times faster

The second result concerns the 7-round version of AES-128. The model proposed a method called Möbius Bridge, which improved the existing meet-in-the-middle attack by 200 to 800 times, depending on the method of calculating execution time. However, it is important to emphasize: this is a purely research scenario. The attack requires a chosen plaintext model with 2^105 queries to the encryptor, making it absolutely impractical for real-world systems.

The full 10-round AES-128 remains secure. This result is a brilliant demonstration of AI's research potential, not a threat to modern cryptography.

Cost of discoveries and new benchmarks

Each of the two studies cost approximately $100,000 in API expenses. The search for the attack on HAWK took about 60 hours, while verification of the AES result took several hundred hours of researcher work. This is comparable to the cost of a small expert team.

To assess the cryptanalytic capabilities of language models, the CryptanalysisBench benchmark was introduced. It includes 191 tasks across six families of cryptographic primitives. Results showed that the best models solve 65–86% of first-level tasks, but on full-fledged schemes without published attacks, their success rate does not exceed 9%.

Analyst's conclusion

We are witnessing a paradigm shift: AI is ceasing to be just a tool and is becoming a full-fledged participant in the research process. For the crypto industry, this means that security audits and algorithm strength analysis can now be conducted faster and more efficiently. However, the main challenge remains the same — verification of results obtained by AI. As the data shows, even advanced models are still far from fully analyzing complex schemes, but the development trajectory is clear and extremely dangerous for outdated or weak cryptographic primitives.