Anthropic's analytical group has presented sensational results: their language model Claude has discovered hidden vulnerabilities in two cryptographic systems that were considered resistant to attacks. These are mathematical algorithms that protect messages, logins, and banking data from outsiders. Most notably, humans had not noticed these flaws for years, while the AI managed it in just a few days.

HAWK: a key that turned out to be weaker than thought

The HAWK system is a digital signature designed with the potential threat of quantum computers in mind. For two years, experts tested its strength and found no problems. However, Claude needed only 60 hours to identify a mathematical pattern that makes HAWK vulnerable. According to Anthropic's calculations, cracking the smallest HAWK key is now possible in 2^38 steps instead of 2^64. This means a reduction in computational cost by approximately 67 million times. There is a solution to the problem, but it is painful: doubling the key size, which negates the algorithm's main advantage.

AES: "Möbius Bridge" accelerates the attack by hundreds of times

As for AES (Advanced Encryption Standard) — the gold standard of encryption that has protected most internet traffic since 2001 — Claude did not crack the full version. The model analyzed a simplified version of the algorithm with 7 rounds instead of the standard 10. The AI found a non-standard approach, called the "Möbius Bridge." The essence of the trick is that one of the selection stages turned out to be redundant. As a result, the most effective attack was accelerated by 200 to 800 times. Notably, Claude worked alone for three days, and then two specialists spent almost a month double-checking its results.

What this means for cryptocurrencies and security

There is no direct threat to Bitcoin (BTC) and Ethereum (ETH) — they operate on different principles, and these attacks do not affect them. However, the news is important for another reason: the machine discovered what leading experts had missed. The main delay now occurs at the stage of verifying results — humans can no longer keep up with the model's pace. This is reminiscent of the story with the SIKE algorithm in 2022, which was cracked on an ordinary laptop in an hour. Time is running: Google plans to transition to post-quantum standards by 2029, and U.S. government agencies have a deadline of 2031.

My expert opinion: This case is a wake-up call for the cryptographic community. We are entering an era where AI can find vulnerabilities faster than humans can verify them. NIST will have to evaluate a public attack on an algorithm that was still being considered for implementation. Anthropic hinted that not all attack results have been published and plans to hold a seminar with scientists. The question is not whether we should trust AI to check security, but whether we can do without it in the future.