Security Current Affairs

Claude Just Broke What Two Years of Experts Missed: Post-Quantum Crypto Gets a Wake-Up Call

2026-07-29

Anthropic's Claude Mythos Preview found a cryptographic weakness in HAWK, a post-quantum signature scheme under consideration for U.S. federal standardization. This is not a drill. The scheme had survived two rounds of expert human review over two years. Claude cracked it in 60 hours.

HAWK is a digital signature system designed to survive future quantum computers. NIST moved it into the third round of its post-quantum signature competition in May, where it's the last lattice-based candidate standing. The AI discovered a mathematical shortcut—a nontrivial automorphism buried in the lattice structure—that nobody had spotted before. For the smallest configuration, the cost of recovering a secret key fell from 2^64 operations to 2^38. That's roughly 67 million times less work.

Here's the rub. Fixing this weakness means roughly doubling HAWK's keys. That eliminates many of the reasons making it attractive in the first place. For blockchain systems evaluating post-quantum replacements, signature size matters acutely—it translates directly to transaction block space and fees. HAWK's compact design was the whole point. That advantage is now gone.

HAWK has never been deployed anywhere. Bitcoin still runs on ECDSA, the pre-quantum signature scheme that candidates like HAWK are eventually meant to replace. But this finding will likely trigger accelerated review within NIST's standardization process and raise uncomfortable questions about whether the agency's timeline accounts for the pace at which AI can surface weaknesses.

The second discovery was equally striking. Claude attacked a 7-round research version of AES-128 and invented a new technique called the Möbius Bridge. It makes an existing attack on a weakened version of AES between 200 and 800 times faster. Notably, Claude initially refused to engage with the problem. It claimed that improving cryptanalysis of AES was impossible because it's the most-studied block cipher in existence. Only after human encouragement did the model proceed to discover the novel technique.

The whole thing cost about $100,000 in API spend. Claude developed and verified the result over approximately 60 hours in a multi-agent environment, with a human researcher providing occasional project-management guidance—but not a lattice-cryptography specialist. Neither flaw affects software now in use. The AES result applies to seven of AES-128's ten rounds and still requires an impractical number of chosen plaintexts. The attack remains exponential. Full-strength AES is secure for production systems.

The verification burden proved heavier than discovery itself. It took just one week for Claude to autonomously discover the improved attack on AES. It took two researchers nearly a month to gain confidence that the method was correct.

Anthropic disclosed both results to the algorithms' authors and to U.S. government and industry partners before publishing. They coordinated the HAWK finding with NIST. The company called the results a substantial research advancement.

As frontier AI models accelerate the discovery of cryptographic weaknesses, the real question surfaces: will human expert review remain the bottleneck—or become the limiting factor—in cryptographic standardization? We're about to find out.

Source & further reading:

Sources