Anthropic Researchers Use Claude Mythos Preview to Find New Cryptographic Attacks on HAWK and Weakened AES

Anthropic researchers report that a Claude model internally referred to as Mythos Preview discovered novel mathematical weaknesses in HAWK, a lattice-based digital signature scheme, and in a reduced-strength version of AES. The findings, detailed in a research write-up and accompanying code repository, represent one of the more concrete public examples of a large language model contributing original results in cryptanalysis rather than merely explaining or applying known techniques.

According to Anthropic, neither result has practical impact on today's computer systems. The AES attack applied to a weakened variant — not full AES-128 or AES-256 as deployed in production systems — and the HAWK finding is a mathematical property rather than an exploitable break of deployed signatures. Anthropic has not disclosed full technical specifics of the attacks in the summary covered here; the repository linked from the research post reportedly contains further detail.

The process took approximately 60 hours of model work and cost an estimated $100,000 in API compute, according to figures shared by Anthropic. Notably, the bulk of human involvement was not technical guidance on the mathematics itself but persistent prompting to keep the model attempting the problem at all. Researchers published excerpts of their prompts, including lines such as "the models tend to think it is impossible to solve so they don't try, they need a good amount of prompting," and "again we are not looking for low hanging fruit, we want proper research to find genuinely hard findings."

Other shared prompts pushed back against the model narrowing its own scope — for instance, refusing suggestions to change the target problem ("no we don't want to change the targets") and insisting on harder variants ("why not do aes-128 r7?"). Researchers also had to repeatedly clarify the actual goal was novel research output, not incremental optimization of known methods: "the whole point is to find something better than existing approaches."

This pattern — a capable model defaulting to conservative, low-effort responses unless explicitly and repeatedly redirected — has been observed in other agentic research contexts, but it is unusual to see it documented this specifically in a hard technical domain like cryptanalysis. The prompts, spelling errors included, suggest the interventions were done in real time under some urgency rather than carefully scripted.

Anthropic has not disclosed pricing, availability, or a public release timeline for Mythos Preview, and it is unclear whether this is a distinct model or a specialized configuration of an existing Claude model. The name suggests an internal research or preview build rather than a shipped product.

What this means

The cryptographic findings themselves are explicitly caveat-laden by Anthropic: no impact on real-world systems. The more significant claim is procedural — that a frontier model, given persistent human prodding, can generate genuinely novel results in a mathematically demanding research field rather than just recombining known literature. The $100,000 cost and 60-hour runtime indicate this is not yet a cheap or fast substitute for human cryptanalysts, and the need for constant human intervention to prevent the model from abandoning difficult subproblems suggests current models still lack the persistence that human researchers bring to open-ended problems by default. Whether this generalizes beyond a curated cryptanalysis setup, or holds up under independent verification of the mathematical claims, remains to be seen since full technical detail has not been widely reviewed outside Anthropic's own repository.