Claude Mythos Weakens HAWK and Speeds Up a Seven-Round AES Attack

Anthropic's HAWK paper describes a previously unused automorphism, or symmetry, in the lattice underlying the signature scheme. The method constructs a public lattice, reduces the key-recovery problem to finding a short vector in roughly half the prior dimension, and reconstructs functionally equivalent signing material.

The released implementation recovers a HAWK-256 secret key end to end within a few hours on a single 96-core server and verifies the result with the HAWK reference implementation. HAWK-256 is a cryptanalytic test parameter, not one of the larger security-level parameter sets. The paper estimates lower attack costs for HAWK-512 and HAWK-1024, but both remain impractical to attack with the published method.

The HAWK team's independent confirmation and withdrawal are more consequential than the challenge-parameter demonstration alone. They show that the result changed the scheme's viability inside an active standards process, while not implying that deployed post-quantum systems were compromised.

The second paper targets seven of AES-128's ten rounds. It removes one guessed key byte from a prior meet-in-the-middle attack through an invariant fingerprint called a Möbius Bridge. Depending on how runtime is counted, Anthropic reports a 200- to 800-fold improvement.

The attack still assumes roughly 2^105 chosen plaintexts and requires infeasible time and memory. The researchers validated components and smaller-scale experiments rather than executing a complete full-cost recovery against seven-round AES-128. Standard ten-round AES-128 remains outside the result, and Anthropic said no production software needs to change.

Anthropic says Mythos Preview produced much of the mathematical work while human researchers supplied direction, compute, and extensive checking. The release includes two technical papers and reproducibility artifacts. The public record also contains confirmation from the HAWK team and NIST's updated status page.

For security and ML practitioners, that sequence is the important one: model-generated hypotheses became actionable only after implementation checks, expert review, public artifacts, and standards-community scrutiny. AI can widen the cryptanalytic search space, but claims this consequential still require evidence that other researchers and institutions can inspect.