Anthropic AI Finds Flaw in Quantum-Resistant Signature Algorithm in 60 Hours

Anthropic’s AI model found a key vulnerability in a next-generation digital signature algorithm in 60 hours, adding a new factor to the cryptocurrency industry’s preparations for quantum computing.

CoinDesk reported on July 30 that Anthropic said its Claude Mythos Preview model discovered an attack technique that halves the effective key strength of HAWK, a candidate quantum-resistant digital signature algorithm, at a computing cost of about $100,000. HAWK had passed two rounds of review over two years by a group of experts.

The research does not directly affect the current security of Bitcoin or Ether. Both networks use elliptic curve signatures, which were not targeted in the attack. HAWK is also not among the algorithms proposed for Bitcoin’s eventual shift to quantum-resistant addresses.

The result is still notable because it underscores how quickly cryptographic attack techniques are advancing. Anthropic said the amount of computation needed to recover a key under HAWK’s minimum parameter set fell to 2^38 from 2^64. BIP-360, a proposal to introduce quantum-resistant addresses for Bitcoin, specifies three algorithms already standardized by the U.S. National Institute of Standards and Technology. It was designed to include multiple algorithms in case one is broken.

Anthropic also disclosed additional findings from the research. The Claude model improved attack efficiency by 200-fold to 800-fold against an intentionally weakened version of the AES algorithm, which is used across the industry to encrypt wallet files. For Poseidon, a hash function widely used in zero-knowledge proof systems, the improvement was less than 10-fold.

Anthropic first notified HAWK’s developers of the vulnerability in June and later disclosed the results after consulting NIST’s public mailing list. Separately, privacy network Zcash activated a new shielded pool upgrade this week designed to remain recoverable in a quantum-computing environment.