Discovering cryptographic weaknesses with Claude
Anthropic ● Covered by 39 sources
Anthropic says its Claude Mythos Preview model found real mathematical flaws in two cryptography systems, not just buggy code. It halved the effective strength of a NIST post-quantum signature contender in just 60 hours of AI work.
Based on reporting by Anthropic — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic has spent the past year showing that Claude can hunt down bugs in cryptographic libraries — the kind of implementation slip-ups that let hackers sneak past encryption meant to protect your bank login. That's useful, but it's also familiar territory for security researchers. What's new, according to a fresh Anthropic writeup, is that Claude Mythos Preview found flaws in the underlying math of the algorithms themselves, not just in how programmers wired them together.
The headline result targets HAWK, a digital signature scheme still competing in NIST's multi-year contest to pick post-quantum cryptography standards. HAWK had already survived two rounds of expert scrutiny over two years. Mythos needed 60 hours to find a previously unexploited symmetry in HAWK's lattice structure, cutting the expected cost of a full key-recovery attack on HAWK-256 from 2^64 down to 2^38. That's not a total break, but it effectively halves the scheme's usable key strength and forces a doubling of key sizes to compensate — which wipes out much of what made HAWK attractive in the first place.
The second result hits closer to home: AES, the encryption standard practically every secure website relies on. Mythos didn't crack full AES-128, but it improved attacks on a weakened, 7-round version researchers use to probe the cipher's margins. Working almost entirely on its own inside a custom scaffold, Claude built something it called a
My take — AI-written commentary, not fact-checked reporting
What sticks with me isn't the math, it's that Claude initially told its handlers the problem was unsolvable — 'AES-128 is just genuinely hard' — and had to be argued into trying harder, which is either a glimpse of appropriate epistemic humility or a preview of AI systems that need to be talked out of giving up. Anthropic grading its own homework here deserves a raised eyebrow, but the multi-agent dynamic where one instance rejected an idea and another ran with it is the actually interesting part, and it's the bit most coverage will skip past in favor of 'AI breaks encryption' headlines it doesn't deserve yet.
Read more about this at: Anthropic