Claude AI Cracks HAWK-256 Post-Quantum Crypto, Speeds Up AES-128 Attack

Anthropic's Claude Mythos Preview has achieved a significant cryptanalytic breakthrough, deriving an end-to-end key-recovery attack against the HAWK-256 post-quantum signature scheme and accelerating an existing seven-round AES-128 attack by 200 to 800 times. The findings, published alongside two technical papers and reproducibility artifacts, were largely conducted by the AI itself, with humans supplying project direction, computing resources, and extensive verification, according to Anthropic. The company stressed that neither result affects production systems, but the demonstration signals a meaningful escalation in AI-assisted cryptanalysis.

The HAWK-256 attack exploits a previously unused symmetry in the lattice structure underpinning the signature scheme. HAWK is the sole lattice-based candidate among nine schemes that NIST advanced to the third round of its additional post-quantum digital-signature process in May 2026, with HAWK-512 and HAWK-1024 as its standard parameter sets—HAWK-256 is a smaller challenge parameter provided as a cryptanalytic target. Researchers Daniël van Gent and Ludo Pulles previously showed that a nontrivial automorphism could reduce HAWK key recovery to finding a short vector in a lattice of roughly half the original dimension, though they noted at the time that it didn't impact HAWK. Anthropic says Mythos Preview found the additional automorphism needed to exploit that path, constructing what the team calls a τ-cocycle lattice from the public key, then using lattice reduction and sieving to recover short vectors before reconstructing a secret basis capable of signing messages for the original public key. The released implementation verifies recovered keys by signing a message and checking it against the NIST reference implementation, yielding a 592-byte decoded key rather than the original 96-byte seed, and supports only HAWK-256—rejecting all other inputs.

The AES-128 result removes a 256-way guessing step from an existing meet-in-the-middle attack, applying to seven of AES-128's ten rounds but still requiring an impractical number of chosen plaintexts. Anthropic's HAWK implementation achieves an expected end-to-end runtime of roughly three hours and 42 minutes on a 96-core server. The repository includes two public keys that Anthropic says it successfully attacked, along with tooling to generate and test fresh HAWK-256 keys. Security teams can audit their own cryptographic exposure by verifying TLS configurations with our SSL/TLS checker and reviewing credential strength using our password checker.

While no production software needs to change today, the results underline both the velocity of AI-driven research and the scrutiny facing post-quantum candidates as NIST narrows its shortlist. Organizations relying on HAWK or AES-128 should track NIST's standardization process, ensure their TLS stacks use modern cipher suites, and run a broader privacy checkup to identify weaknesses before adversaries do.