Claude AI Just Cracked a Post-Quantum Test Scheme and Found a Faster 7-Round AES Attack
20:59 · July 28, 2026 · Hacker News AI Section
Anthropic says Claude Mythos Preview helped derive an end-to-end key-recovery attack against HAWK-256 and a 200- to 800-fold speedup for an attack on seven-round AES-128. The HAWK attack exploits a previously unused symmetry in the lattice behind the signature scheme. Anthropic's released implementation gives an expected end-to-end runtime of about three hours and 42 minutes on a 96-core server
Summary
Anthropic reports that its Claude Mythos Preview model contributed to an end-to-end key-recovery attack on the HAWK-256 parameter set of the lattice-based signature scheme under review by NIST. The attack exploits an additional automorphism that reduces the search-module Lattice Isomorphism Problem to finding a short vector in a lattice of roughly half the original dimension. The resulting procedure builds a τ-cocycle lattice from the public key, applies lattice reduction and sieving, and reconstructs a functionally equivalent 592-byte signing key. The released implementation runs in an expected three hours and 42 minutes on a 96-core server, lowering the estimated work factor from 2^64 to 2^38 for this challenge parameter only.
The same model produced a second result that removes a 256-way guessing step from an existing meet-in-the-middle attack on seven-round AES-128. It does so by identifying an invariant the researchers term the Möbius Bridge, which remains constant across the enumerated values and therefore allows the enumeration to be dropped. After accounting for the cost of the transform and other adjustments, the projected speedup ranges between 200- and 800-fold, depending on the chosen runtime metric. The attack still requires approximately 2^105 chosen plaintexts and applies only to the reduced-round construction.
Anthropic states that neither finding affects production systems. HAWK-256 is supplied solely as a cryptanalytic target; the two NIST security-level parameter sets remain outside the demonstrated reach of the method. Full-round AES-128 is likewise untouched. The company released two technical papers, reproducibility artifacts, and code that supports only HAWK-256 and verifies recovered keys against the NIST reference implementation. The model operated for roughly 60 hours in a multi-agent environment at an API cost of about $100,000, while human researchers spent several hundred hours on verification. The attack remains exponential and does not extend to other lattice-based candidates.
Why it matters
Directly demonstrates AI-assisted cryptanalysis on post-quantum signatures and reduced-round AES, actionable for Dutch/EU security teams evaluating NIST candidates and AI risks in crypto. Addresses security vulnerabilities with concrete attack paths and mitigations.
