Anthropic’s Frontier Red Team has disclosed research showing that its internal model, Claude Mythos Preview, found two cryptographic weaknesses. The company was explicit about the limit of those results: neither one breaks encryption that people and institutions use in the real world today.
A weakness found in the HAWK post-quantum candidate
The first result involves HAWK, a post-quantum cryptography digital signature candidate currently under review by the U.S. National Institute of Standards and Technology, or NIST.
According to Anthropic, Claude Mythos Preview identified a previously unknown mathematical symmetry in HAWK’s lattice structure, technically described as a non-trivial automorphism. Based on that, the model proposed an improved attack that cut the effective key strength in half. For HAWK-256, Anthropic said the attack cost fell from 2^64 to 2^38.
The model carried out that work over roughly 60 hours, using about $100,000 worth of API compute. Anthropic said the attack remains theoretical because HAWK is only a candidate scheme and has not been deployed in practice.
Reduced-round AES-128 attacked faster
The second result targeted a 7-round version of AES-128. Full AES uses 10 rounds. Anthropic said the model developed a fingerprinting algorithm called Möbius Bridge that removes one enumeration step, making the attack run 200 to 800 times faster than earlier meet-in-the-middle methods.
That result also came out of a largely autonomous process. Anthropic said the model generated hypotheses, ran experiments, and repeatedly revised its approach over about three days, producing billions of tokens before reaching the final result. Human involvement was limited to minimal guidance.
Anthropic added that this attack applies only to a reduced-round, deliberately weakened AES variant, not to full AES. In the company’s description, it does not create a practical threat to deployed encryption.
The main takeaway is about AI research capability
Anthropic said the significance of the work lies in the research process itself. The company built an environment for Claude that let the model form hypotheses, run experiments, and refine ideas step by step, so it could execute a full cryptanalysis workflow with humans focused mainly on verification and project management rather than technical direction.
In Anthropic’s framing, the results show that AI can do cryptanalysis research. They do not show that Claude has broken modern encryption systems. The two targets were an unadopted HAWK candidate and a weakened reduced-round AES variant, which means real-world security is not affected.
Anthropic also highlighted one detail from the project: human researchers spent more time verifying the model’s mathematical claims than the AI spent producing the findings.