Claude Breaks Post-Quantum Crypto: HAWK-256 Key Strength Halved

Anthropic's Claude Mythos Preview discovered a practical key-recovery attack on HAWK-256, a NIST post-quantum candidate, halving its effective security in ~60 hours of autonomous work — and found a new attack vector against round-reduced AES.

Anthropic researchers used Claude Mythos Preview to discover improved attacks against two major cryptographic algorithms — including a NIST post-quantum candidate that had survived two years of expert human review.

HAWK-256: Key strength cut in half. HAWK is a lattice-based digital signature scheme under consideration by NIST for the post-quantum cryptography standard. Despite passing two rounds of scrutiny, Claude found a practical key-recovery attack in approximately 60 hours of autonomous work — effectively halving HAWK's security margin. The attack has been disclosed to the HAWK team and NIST.

AES: New angle on the world's most-used cipher. Claude also identified a novel attack vector against round-reduced versions of the Advanced Encryption Standard (AES), the symmetric cipher protecting virtually all encrypted internet traffic since its adoption in 2001. While the attack targets reduced-round variants (not full AES), it represents a fresh approach that had eluded human cryptanalysts.

Critically, neither finding threatens any production system today. HAWK is still a candidate, not yet deployed. The round-reduced AES attack does not apply to the full 10/12/14-round standard. But the implications are significant: this is the first time an AI has found mathematical flaws in cryptographic algorithms themselves — not just bugs in how code implements them.

The Frontier Red Team at Anthropic framed the work as a proof-of-concept for AI-assisted cryptanalysis, suggesting that models like Claude could accelerate the discovery of weaknesses before adversaries do.

Sources: Anthropic Research, GitHub: cryptography-research-demo, NIST HAWK Specification