Artificial intelligence continues to surprise even the most experienced cryptographers. This time, Anthropic's Claude model discovered critical vulnerabilities in two encryption systems that were considered secure. These are mathematical algorithms underlying the protection of messages, logins, and banking data from external attacks. And although current systems remain safe, the very fact that a machine identified weaknesses faster and more efficiently than leading experts deserves close attention.
What's wrong with HAWK and AES?
HAWK is a digital signature specifically designed to protect against quantum threats. For two years, experts could not find any flaws in it. However, Claude needed only 60 hours to identify a hidden pattern in its mathematical foundation. As a result, according to Anthropic's calculations, cracking the smallest HAWK key is now possible in 2^38 steps instead of 2^64, reducing computational complexity by approximately 67 million times. There is a solution — doubling the key size, but this completely negates the algorithm's main advantage.
In the case of AES (Advanced Encryption Standard), which has protected the bulk of internet traffic since 2001, Claude did not crack the full version. Instead, it analyzed a simplified version with 7 rounds instead of 10. Here, the model found a non-standard attack method called Möbius Bridge. The essence is that one of the selection stages turned out to be redundant, allowing the attack to be accelerated by 200–800 times. Claude took three days to work independently, after which two specialists spent nearly a month double-checking the result.
Why this changes the game
The main takeaway from this research is that humans can no longer keep up with the pace of AI work. Verifying results has become a bottleneck. This is not the first time algorithms considered secure have proven vulnerable. In 2022, researchers cracked another post-quantum algorithm, SIKE, in about an hour on a regular laptop.
For cryptocurrencies such as Bitcoin (BTC) and Ethereum (ETH), these attacks do not yet pose a direct threat — they operate on different principles. However, time is running out. Google plans to transition to new standards by 2029, and U.S. government agencies have been given a "quantum deadline" of 2031.
Now, NIST will need to evaluate the public attack on an algorithm that has not yet been implemented. Anthropic has indicated that not all results have been published yet. In the near future, the company will hold a special seminar with scientists.
Expert opinion: This work is not just a technical discovery. It calls into question the very paradigm of manual verification of cryptographic protocols. If AI can find in days what has eluded humans for years, then the future of cybersecurity will depend not on the strength of algorithms, but on the speed of their automated auditing. The cryptocurrency market should prepare for the possibility that the "quantum threat" may not be the only one — AI attacks could become a new reality in the coming years.