Anthropic's Claude Mythos Cracks Cryptographic Defenses That Held Up for Years

A model called Claude Mythos Preview spent 60 hours and roughly $100,000 in compute finding a stronger attack against HAWK, a post-quantum signature scheme still competing in NIST's third-round standardization process. Human cryptographers had scrutinized HAWK for more than two years without surfacing the flaw Mythos exposed.

The weakness lies in a previously unnoticed symmetry inside the mathematical lattice that underpins HAWK's security guarantees. Anthropic disclosed the finding alongside a second result: a new attack on AES-128 reduced to seven rounds, built around a technique the model itself named "Möbius Bridge."

That method eliminates one of the guesses an attacker needs to make, improving on the best known prior attacks by a factor of somewhere between 200 and 800. Anthropic was clear that neither result threatens systems currently in production. Full AES-128 runs 10 rounds, not seven, and HAWK remains a research candidate rather than a deployed standard.

A model that initially said no

Getting there wasn't straightforward. When first asked to attack AES, Mythos pushed back, telling researchers plainly that "AES-128 r5/r6 is just genuinely hard" and that a different outcome would require changing the target.

Researchers responded by building a scaffold that let the model form its own hypotheses and test them experimentally, rather than following a fixed attack template. That looser structure proved decisive: Mythos went on to generate several hundred million tokens across three days of work on the AES problem, guided by only three substantive human prompts, at a cost of about $100,000 for roughly a billion tokens total.

The HAWK breakthrough followed a similar pattern of machine-driven persistence overriding early skepticism, including its own. Mythos operated inside a multi-agent setup where one instance initially dismissed the winning attack idea as infeasible, only for a second agent to push through and fully exploit it.

The human researcher on the HAWK project had a background in theoretical computer science but no specialist expertise in lattice-based cryptography, and mostly handled project management rather than the mathematics itself.

That detail matters more than the headline numbers. It suggests the bottleneck in this kind of cryptanalysis is shifting away from deep domain specialization and toward the ability to structure a problem so an AI system can explore it.

Anthropic's framing throughout is cautious rather than alarmist: these are advances against reduced or candidate schemes, not live infrastructure. But a system that outpaces two years of expert review on a NIST finalist, with a non-specialist largely just steering the process, is a preview of how AI-assisted red-teaming could reshape cryptographic standardization timelines going forward.