Artificial intelligence firm Anthropic announced this week that its Claude Mythos Preview model discovered an attack method in roughly 60 hours, effectively halving the key strength of the HAWK algorithm, a candidate designed to replace current digital signatures. The attack cost approximately $100,000 in computing resources, and HAWK had not yet been publicly deployed in any system after undergoing two years of expert review across two rounds. HAWK is one of the post-quantum digital signature candidates under scrutiny, intended to replace the elliptic curve signature algorithms currently securing online banking and web payments.

Anthropic's research shows that for HAWK's smallest parameter set, the expected computational effort to recover a key dropped from roughly 2^64 operations to 2^38. While larger keys remain difficult to crack, doubling the key length to compensate for the vulnerability would significantly erode HAWK's performance advantages, making it far less competitive. The company disclosed the attack to HAWK's authors in June and coordinated the release of findings with the National Institute of Standards and Technology (NIST).

This research has no direct impact on Bitcoin or Ethereum, as both currently rely on elliptic curve signatures for transaction security, and HAWK is not among the algorithms targeted for Bitcoin's migration plan. According to Bitcoin Improvement Proposal BIP-360, quantum-resistant addresses designed to address quantum threats will use three algorithms already standardized by NIST, with multiple backup options in case some algorithms are broken by future technological advances. Meanwhile, the companion proposal BIP-361 notes that the migration window is narrowing, as cryptographic attack methods improve at a rate of up to 20 times.

Anthropic's latest results confirm this trend. In another experiment, the model achieved a 200 to 800 times improvement in attack efficiency against a weakened version of the AES algorithm, widely used in the industry for encrypting wallet files. Notably, the Claude model initially refused to process the problem, describing it as "truly difficult" with "no easy opening," but after three short prompts from researchers, the model generated one billion output tokens over three days and eventually found an improved method. Verification of the findings required nearly one month of work from two researchers.

For the Poseidon hash function, which underpins zero-knowledge proof systems including rollups and privacy protocols, the model's improvements were limited to no more than 10 times, resulting in a relatively minor impact. The timing of this progress coincides closely with a technical upgrade from the privacy network Zcash. Zcash activated an upgrade on Tuesday, introducing a new shielded pool designed to ensure asset recoverability even when quantum computers become available.

Currently, various cryptographic schemes aimed at helping crypto assets transition to quantum resistance are facing continuous stress tests from AI systems that work faster than human designers.