Claude Mythos Cracks HAWK-256, Speeds AES Attack

Anthropic published results this week showing its Claude Mythos Preview AI model derived two things: a complete, end-to-end key-recovery attack against HAWK-256, a lattice-based post-quantum signature scheme, and a 200-to-800-fold speedup for an existing attack on seven-round AES-128.

The HAWK-256 result is the more significant of the two. The attack exploits what Anthropic describes as a previously unidentified symmetry in the lattice structure underlying the scheme. Anthropic’s implementation runs end-to-end in approximately three hours and 42 minutes on a 96-core server — not fast enough for real-time exploitation, but fast enough to be a result that HAWK-256’s designers need to address. Source: The Hacker News, citing Anthropic’s own release.

The AES-128 result is narrower. Seven-round AES-128 is a reduced-round construction used in academic cryptanalysis benchmarking — not the full ten-round AES-128 in production use. A 200–800× speedup on a reduced variant is a meaningful academic finding. It does not mean production AES-128 is broken.

What actually matters here

Post-quantum migration is already underway. NIST finalized its first post-quantum standards last year — ML-KEM, ML-DSA, SLH-DSA, and LMS — and the practical work of replacing RSA and elliptic-curve systems has started, if slowly, across most regulated sectors. HAWK-256 is not among those four finalists, so this result doesn’t destabilize the primary migration path.

What it does is demonstrate that AI-assisted cryptanalysis is producing research-grade outputs. That’s the part worth watching.

Cryptanalysis has historically been slow, specialized work — small teams, years of iteration, the occasional lucky insight. If AI models can materially compress the cycle from “underexplored structural property” to “working attack,” the academic cryptographic community’s workload just changed. Not just on post-quantum candidates. On everything in the research pipeline.

For practitioners, the near-term read is fairly stable: NIST’s selected post-quantum standards remain sound based on current research, and the transition guidance hasn’t changed. Organizations relying on post-quantum candidates outside the four finalized standards should pay closer attention to the cryptanalysis literature than they did a year ago.

The longer-term question is what AI-assisted cryptanalysis looks like in three years, when the models are better and the target pool is larger. That’s worth keeping an eye on — not as a reason to stop migrating, but as a reason to stay close to NIST’s current guidance rather than improvising.