Anthropic’s Frontier Red Team published research on July 28 showing that its unreleased Claude Mythos Preview model independently produced two novel cryptanalytic results: a stronger key-recovery attack against HAWK, a NIST post-quantum digital signature candidate, and a substantially faster attack against a reduced, seven-round research version of AES-128. Anthropic frames the work as evidence that frontier models can now contribute original results in a field long considered one of the hardest to automate, while stressing that neither attack has any practical impact on deployed software.
What Mythos found
Working largely autonomously in a multi-agent setup — with human researchers mostly limited to project management, simple prompting, and later verification — Mythos identified a previously unexploited mathematical symmetry (an automorphism) in HAWK’s lattice structure. That let it demonstrate full key recovery on small HAWK-256 parameters at roughly 2^38 operations, far below the roughly 2^64 previously assumed, effectively halving the scheme’s expected key strength despite HAWK having already survived two rounds of expert human cryptographic review over two years. On AES, Mythos developed a new technique researchers are calling “Mobius Bridge” that improves the best-known attacks on seven-round AES-128 by a factor of 200 to 800, compared to prior published results. Anthropic said each result cost on the order of $100,000 in API compute to produce.
Why it doesn’t affect real systems
Both results come with hard limits on real-world relevance. HAWK is only a candidate post-quantum signature scheme still under NIST evaluation and is not deployed anywhere in production. The AES attack targets a deliberately weakened seven-round research variant, not the full ten-round AES-128 used across the internet — full AES remains unbroken. Anthropic said no production software needs to change as a result of either finding, and independent cryptographers who reviewed the work broadly agreed the results are genuine but narrow, improving the field’s understanding of security margins rather than exposing a live vulnerability.
Why it matters
The research lands as frontier labs increasingly promote their models as tools for scientific discovery rather than just chat assistants, and cryptanalysis has historically been a poor fit for AI automation because it requires long chains of exact, verifiable mathematical reasoning rather than pattern matching. Anthropic’s own framing is cautious: the company has previously flagged AI-assisted cryptanalysis as a capability worth monitoring for both defensive and offensive implications, and this result — a real if narrow contribution against a non-deployed scheme and a weakened cipher variant — is likely to be read as an early, contained data point in that broader safety conversation rather than a signal that deployed encryption is newly at risk.