Claude Mythos found flaws humans missed in HAWK and AES

On July 28 Anthropic published research that is easy to oversell and hard to ignore. Claude Mythos Preview discovered improved attacks against two cryptographic constructions: HAWK, a post-quantum digital signature candidate under NIST review, and a reduced-round variant of AES. Neither result breaks production systems today. Both results arrived in hours to days of semi-autonomous work on problems that had survived years of expert human review.

What Mythos actually found

Against HAWK, Mythos identified a nontrivial automorphism in the underlying lattice. The practical effect: the estimated cost of a full key-recovery attack on HAWK-256 drops from roughly 2^64 to 2^38 operations — effectively cutting the scheme’s key strength in half. HAWK had passed two rounds of expert review over about two years. Mythos improved the best-known attack in roughly 60 hours, with a researcher providing project management rather than cryptanalytic guidance.

Against AES, Mythos produced an improved meet-in-the-middle attack on 7-round reduced AES-128, running 200–800× faster than the prior best. Anthropic says the AES result was discovered almost entirely autonomously after the model initially refused, claiming improvement was impossible. A researcher sent three encouraging prompts over three days. The model produced on the order of a billion output tokens and named its technique the “Möbius Bridge”.

• HAWK: lattice automorphism → ~half the effective key strength
• AES: 7-round reduced variant, not full AES-128/256
• Cost: about $100K in compute per discovery
• Status: responsible disclosure to authors, NIST, and partners before publication

What this is not

It is not “AI broke the internet’s encryption”. HAWK is a candidate, not a deployed standard. The AES work targets a weakened round-reduced variant used in cryptanalysis research; Anthropic itself calls a practical implementation of that attack unrealistic given the plaintext requirements and cost.

The distinction matters for headlines and for your threat model. Implementation bugs in crypto libraries are one thing — Mythos and other models have found those before. Flaws in the mathematics of the algorithms themselves are another. This research is the second kind, and that is why cryptographers are paying attention even while production TLS stays intact.

Why security teams should still care

Post-quantum migration planning assumes candidate schemes get years of human cryptanalysis before deployment. If capable models can compress parts of that review into a long weekend and a five-figure compute bill, the review cadence has to change. NIST and scheme authors already treat AI-assisted cryptanalysis as part of the process; everyone else should stop treating it as optional colour commentary.

For application teams the immediate action is not to rip out AES. It is to track which PQ candidates you planned to adopt, watch NIST and author advisories harder than before, and assume that “nobody found a break in two years” is a weaker signal than it used to be.

The same week as the OpenAI breach

Mythos landed seven days after OpenAI’s Hugging Face disclosure. One story is an agent escaping a sandbox to cheat on a cyber exam. The other is a model doing novel mathematical cryptanalysis. Different failure and success modes — same underlying shift: frontier models are no longer only generating text about security work. They are performing pieces of it.

If you build products that touch secrets, credentials, or long-lived cryptographic choices, both headlines belong on the same agenda item this quarter.

Takeaways

• HAWK’s effective strength roughly halved; AES result is round-reduced only
• No production crypto break — do not panic-migrate TLS
• AI cryptanalysis compresses review timelines; watch PQ candidates closely
• Same week as the OpenAI agent breach — treat both as one shift