Anthropic's Claude Mythos found real weaknesses in expert-reviewed encryption

AI just did something cryptographers spent years failing to do.

On July 28, 2026, Anthropic published research showing its unreleased Claude Mythos model found real mathematical weaknesses in two encryption systems that had already survived expert review.

The two results

The first target was HAWK, a post-quantum signature scheme that had held up through two years and two rounds of expert review. Claude Mythos cut its effective key strength in half in about 60 hours, dropping the expected attack cost on HAWK-256 from 2^64 to 2^38 operations.

The second was a 7-round version of AES. The model found a shortcut, nicknamed the “Möbius Bridge,” that sped up the best known attack by 200 to 800 times, beating a record set by human cryptographers in 2013.

Neither result came cheap or unverified. Each cost roughly $100,000 in compute, and Anthropic staff spent weeks verifying the AES work was real.

What is not at risk

Two caveats matter before anyone panics about their bank account. Nothing you use today is threatened by these results. HAWK is not deployed anywhere, and standard AES-128 uses 10 rounds, not the 7-round reduced version the model attacked.

This is a research milestone, not a broken lock. Reduced-round attacks and pre-deployment schemes are exactly where cryptanalysis progress has always shown up first, whether the analyst is human or machine.

Why it still matters

The uncomfortable part is what the results say about the review process itself. HAWK’s credential was precisely that experts had studied it for two years and found nothing. An AI model then found a weakness in 60 hours.

AI is starting to out-search humans in the exact domains where “years of expert review” used to be the safety guarantee. For post-quantum cryptography, which is being standardized right now on the strength of such reviews, that changes the calculus: the same models that threaten schemes can also harden them, if they are pointed at the problem before deployment rather than after.