A research team from Anthropic has announced a significant breakthrough in cryptanalysis. Their language model, Claude, operating in Mythos Preview mode, managed to identify previously unknown vulnerabilities in two key encryption systems: the post-quantum signature scheme HAWK and a simplified version of the AES standard. The result is notable not so much for the flaws themselves, but for the speed and methodology of their discovery.

HAWK: 67 Million Times Faster — At What Cost?

HAWK is a promising digital signature system designed to defend against attacks using quantum computers. For two years, the brightest human minds could not find any flaws in it. However, Claude succeeded in 60 hours. It uncovered a hidden mathematical pattern that allows the smallest HAWK key to be recovered not in 2^64 steps, as assumed, but in just 2^38 steps. This reduces the computational complexity by approximately 67 million times. A solution exists, but it is painful: the key size must be at least doubled, effectively negating one of the scheme's main advantages—its compactness.

AES and the "Möbius Bridge"

In the case of AES, which has protected the vast majority of internet traffic since 2001, Claude attacked not the full algorithm, but a reduced version with 7 rounds instead of the standard 10. The model developed an unconventional attack, dubbed the "Möbius Bridge." The trick lies in the fact that one of the key brute-force stages turned out to be redundant. As a result, the most effective attack found was accelerated by a factor of 200 to 800. Claude completed the entire task in three days. Afterward, two Anthropic specialists spent nearly a month verifying and confirming its conclusions.

A New Reality: AI as the Chief Security Auditor

This event changes the rules of the game. Previously, we relied on the slow but meticulous work of expert cryptographers. Now it is clear: the human brain can no longer keep pace with the tempo set by neural networks. The primary bottleneck becomes the stage of verifying results produced by AI. This is not the first time mathematical elegance has let developers down—in 2022, the post-quantum algorithm SIKE was cracked on an ordinary laptop in an hour. But the current issue is that AI finds vulnerabilities faster than humans can verify them. NIST will now have to evaluate a public attack on an algorithm that was under consideration for adoption. Anthropic hints that not all results have been published yet and is preparing a special seminar for the scientific community.

Expert Opinion: This is a wake-up call for the crypto industry. Bitcoin and Ethereum are not yet in danger—their algorithms remain unaffected—but the question of "Q-Day" is becoming increasingly real. Google plans to migrate to post-quantum standards by 2029, while U.S. government agencies have a deadline of 2031. If AI can already find such "holes" in days, humanity will have to drastically accelerate the transition to new cryptography. Should the security of your data be entrusted to artificial intelligence? It seems we may have no choice left.