Bitcoin's long-term plan for surviving quantum computers rests on an assumption that some of the algorithms being considered as replacements will eventually be broken — just not the ones Bitcoin actually picks. That assumption got a real-world test this week after an AI model weakened one candidate algorithm far faster than cryptographers expected.
According to CoinDesk, Anthropic's Claude Mythos Preview model found a vulnerability in HAWK, a proposed digital-signature scheme designed as a replacement for the signatures currently securing web traffic and banking systems. The attack cut the security of HAWK's smallest parameter set roughly in half, reducing the number of operations needed to break it from about 2^64 down to around 2^38.
A Fast, Cheap Attack on a Well-Reviewed Algorithm
The result is notable less for the algorithm itself — HAWK isn't deployed anywhere in production, and Bitcoin does not use it — than for how quickly and cheaply it came together. The AI model needed about 60 hours of computing time and roughly $100,000 to produce the attack, compared with two full years and two separate rounds of expert human review that HAWK had already been through before this.
Researchers found the flaw in June 2026 and disclosed it privately to HAWK's authors before CoinDesk's July 30 report. Larger HAWK key sizes remain impractical to attack with the same method, but the need to compensate with bigger keys makes the algorithm considerably less attractive as a lightweight signature standard going forward.
Why Bitcoin Cares About an Algorithm It Doesn't Use
Bitcoin and Ethereum signatures today rely on elliptic curve cryptography, which is unaffected by this specific attack. But Bitcoin's proposed quantum-resistant migration path, outlined in BIP-360, specifies three NIST-standardized algorithms rather than HAWK — and the broader logic behind that proposal, laid out in a companion document known as BIP-361, is that the safe migration window is closing faster than expected as quantum and classical attacks alike keep improving, in some cases by as much as twentyfold.
The HAWK result adds a concrete data point to that argument: an algorithm that survived two years of expert scrutiny lost half its effective security to an AI system in less than three days of compute time.
Other Algorithms Took Hits Too
The same research effort reported improvements against other cryptographic targets. Attacks against the AES cipher improved by factors of roughly 200 to 800, and attacks against the Poseidon hash function — which underpins many zero-knowledge proofs and blockchain rollups — improved by just under tenfold. Notably, the model reportedly refused to work on the AES problem at first, only producing results after three separate prompts and generating roughly a billion output tokens over three days of work; verifying those results then took two human researchers nearly a month.
The timing overlaps with Zcash's activation, on Tuesday, of a network upgrade introducing a quantum-resistant shielded pool — one of the first live deployments of post-quantum cryptography in a major blockchain, and a sign that some projects are no longer waiting for the debate over which algorithms to trust before starting to migrate.
