Artificial intelligence continues to demonstrate that its potential extends far beyond text generation. In its own research, Anthropic's Claude Mythos Preview model identified two previously unknown cryptanalytic attacks. The first targets the post-quantum digital signature scheme HAWK, and the second targets a reduced 7-round version of the AES-128 cipher.

Although both findings are primarily research-oriented and do not threaten current industrial systems, they are a significant indicator. We are witnessing advanced AI models beginning to operate at the level of expert cryptographers, which opens up both new opportunities for security auditing and potential risks in the future.

Attack on HAWK: Undermining Trust in Post-Quantum Signatures

HAWK is one of the candidates for standardizing post-quantum digital signatures, designed to protect against attacks using quantum computers. Claude Mythos Preview discovered a hidden symmetry in the mathematical structure of HAWK, which significantly accelerated the private key recovery attack.

For the HAWK-256 version, the cost of full key recovery decreased from 264 to 238 operations. This means that to maintain the previous level of security, developers would have to roughly double the key sizes, negating one of the scheme's main advantages — its compactness. It is important to emphasize: the attack does not affect the HAWK-512 and HAWK-1024 versions, nor does it extend to other post-quantum lattice-based schemes.

Möbius Bridge: Accelerating the Attack on 7-Round AES

The second finding concerns the 7-round version of AES-128 — a reduced academic variant used for security testing. Claude Mythos Preview developed a method called Möbius Bridge, which improved the best-known meet-in-the-middle attack. The acceleration ranged from 200 to 800 times, depending on the calculation methodology.

However, this scenario remains purely theoretical. It assumes a chosen plaintext model where the attacker can send 2105 pre-selected encryption queries. The full 10-round AES-128 used in real-world systems remains unshaken.

A New Benchmark for Evaluating AI Cryptanalysts

As part of the research, Anthropic, together with scientists from ETH Zurich, Tel Aviv University, and other institutions, presented CryptanalysisBench — a specialized benchmark for evaluating the cryptanalytic capabilities of language models. It includes 191 tasks across six families of cryptographic primitives.

The results showed that the best models, including Claude Mythos 5, solve 85-86% of first-level tasks. However, on full, unweakened schemes, no model exceeded 9%. This clearly demonstrates that we are still far from creating a universal AI-based "cipher breaker."

Expert Opinion: Anthropic's research is not a cause for panic, but an important step forward in the automation of cryptanalysis. Such tools in the hands of "white hat" hackers and standards developers can significantly accelerate the identification of weaknesses at the design stage. However, the flip side is the potential use of such models by malicious actors. The industry needs to prepare for a new reality where AI becomes a standard tool for both defense and attack.