Anthropic says Claude Mythos found new weaknesses in post-quantum and reduced-round AES

Anthropic said its Claude Mythos Preview model, working with the company’s Frontier Red Team, developed an attack that cuts the cost of recovering a HAWK-256 key from about 2^64 to 2^38 operations. The research, published July 28, 2026, describes an AI model producing genuinely novel cryptanalysis rather than reproducing known results.

HAWK is a post-quantum digital-signature candidate under review at the National Institute of Standards and Technology, or NIST, and had survived roughly two years of expert human scrutiny. Anthropic said the model found its improved attack in about 60 hours. Against a 7-round reduced version of AES-128, it built a meet-in-the-middle attack over about three days – generating roughly 1 billion tokens – that runs 200 to 800 times faster than the best previously known method, using a technique the company calls a “Mobius Bridge.”

The caveats are large, and Anthropic states them. HAWK is not deployed anywhere. The AES result targets a deliberately weakened 7-round variant and does not touch the full 10-round AES-128 that protects production systems today. Each result cost about $100,000 in API compute to produce, and the work leaned on CryptanalysisBench, an evaluation built with ETH Zurich, Tel Aviv University and TU Berlin.

The findings are also a vendor’s own account of its unreleased model, not independently reproduced by outside cryptographers. Novel attacks on reduced-round ciphers and undeployed candidates are how the field stress-tests designs; they are not breaks of live encryption.

Whether outside researchers can reproduce the HAWK-256 attack, and whether NIST factors it into its post-quantum decisions, will show how much the result matters.