On July 28, Anthropic’s Frontier Red Team published a result that genuinely impressed me — and left me a little uneasy. Using Claude Mythos Preview, researchers didn’t just find bugs in crypto software; they found mathematical weaknesses in the algorithms themselves. That’s a qualitative leap.
From implementation bug to breaking the scheme
Until now the story was: Mythos is good at finding implementation errors — the spots where programmers wired up encryption incorrectly. That’s dangerous, but fixable. This goes a level deeper. Mythos found ways to attack the underlying math.
Two results
The first concerns HAWK, a digital signature scheme built for a world with quantum computers. HAWK is one of the third-round candidates from NIST’s 2022 call for post-quantum schemes. Despite two years and two rounds of review by human experts, Mythos improved the best-known attack in just 60 hours — effectively cutting its key strength in half.
The second is a new attack on round-reduced AES, the reduced-round variant of the world’s most widely used symmetric cipher. For the HAWK result a researcher worked alongside Claude; for the AES attack another built a scaffold that let Claude discover it fully autonomously. Each run cost roughly $100,000 in API compute.
Crucially, and Anthropic stresses this clearly: both attacks are substantial research advances, but not a threat to production systems. Your bank and your email are not at risk today.
CryptanalysisBench
A new benchmark came out of the work, too. Together with researchers at ETH Zurich, Tel Aviv University and the University of Haifa, Anthropic built CryptanalysisBench — a collection of ciphers that makes it easy to measure the cryptanalytic ability of LLMs.
My take
The most telling detail is buried in the process: Simon Willison highlighted the shared prompts, in which the model had to be repeatedly encouraged not to give up and to “find something that’s worth publishing.” So AI-assisted cryptanalysis isn’t a one-click affair yet — but it works. And it cuts both ways: defenders can harden schemes earlier, while attackers gain a powerful new tool. For post-quantum standardization it’s a wake-up call: if a model can do in 60 hours what two years of expert review missed, every candidate should now go through the AI grinder too.
Sources: