Anthropic's Claude AI has discovered a serious weakness in the HAWK digital signature scheme, a candidate for post-quantum cryptography still under evaluation by NIST. The attack reduces the complexity of breaking HAWK from 2^64 to 2^38 steps — a 67 million-fold improvement. The fix requires doubling the key size, which undermines the scheme's main selling point: efficiency.
60 hours vs. two years
Claude worked on HAWK for 60 hours before finding the vulnerability. Human experts had reviewed the same scheme for two years without spotting it. The finding is a stark reminder of how AI can outpace human cryptanalysis. It also puts NIST in a bind: the agency must now decide whether to continue evaluating a scheme that has a known public attack.
The Möbius Bridge attack
Claude also found a new attack on a reduced-round version of AES — 7 out of 10 rounds — using a technique called the Möbius Bridge. The attack is 200 to 800 times faster than previous methods. Human researchers took nearly a month to verify Claude's AES findings. That's a long time, but it's still faster than the two years it took to miss the HAWK flaw.
Bitcoin and Ethereum safe
The attacks don't affect Bitcoin or Ethereum, which rely on different cryptographic systems. But the broader push to quantum-resistant crypto is real. Google plans to switch by 2029; US agencies face a 2031 deadline. The HAWK finding could force a rethink of which schemes are ready for prime time.
Anthropic's next move
Anthropic has hinted at additional unreleased results and plans a workshop with academics. The company is clearly positioning Claude as a tool for serious cryptanalysis. For NIST, the clock is ticking. The agency has to decide whether HAWK can be salvaged — or whether to drop it entirely.
