AI Security Series · Part 12
Last time, the theme was: the tools got brilliant, the bugs stayed boring — state-of-the-art AI pointed at the same dumb, preventable mistakes that have always breached companies. This entry is the other end of that spectrum. This time the tooling did something that isn't boring at all.
On July 28, Anthropic published research in which Claude produced genuinely novel cryptanalysis — original mathematics — against real cryptographic schemes. And here is the part every other headline is going to bury: nothing anyone uses is broken. Both halves of that sentence are the story. Skip either one and you've got it wrong.
What actually happened
The setup was Claude Mythos Preview running as multiple collaborating agents in a sandboxed, Claude Code–like harness with access to computational tools (Python, Sage) and the published cryptographic literature. Two results came out of it.
1. A new symmetry in HAWK
HAWK is a post-quantum digital signature scheme — one of the candidates under NIST's third-round consideration for additional signature standards. Claude found a nontrivial automorphism (a hidden symmetry) in the lattice HAWK is built on. Prior work had proven that if such an automorphism could be found efficiently it would enable an attack — but nobody had established that it was actually reachable. Claude reached it.
The effect: for HAWK-256, the expected cryptanalysis cost drops from roughly 264 to 238 — effectively halving the key strength. It took about 60 hours of work (and, per Anthropic, roughly $100,000 in API compute).
2. A faster attack on 7-round AES
The second result is a fingerprinting technique Anthropic calls the "Möbius Bridge," which improves prior meet-in-the-middle attacks by eliminating an enumeration step — a 200–800× speedup, a factor-of-256 cut in required lookups. This one took roughly one billion output tokens over three days of autonomous work.
Read the next sentence carefully, because it's the one that matters: this attack is on a reduced-round variant — 7 of AES's 10 rounds. Full AES-128 is not affected and remains secure. In Anthropic's own words, the attack "does not break the full cipher," and even against the reduced version it "would cost hundreds of millions of dollars to implement." If you use AES — and you almost certainly do, every time you load an HTTPS page — nothing changes.
Why "nothing is broken" is the point, not a disclaimer
By the time you read this, plenty of coverage will say some version of "AI breaks encryption." That framing is simply false, and refusing it is exactly why this piece is worth writing.
- HAWK is not deployed anywhere. It's a candidate, not a shipped standard. As Anthropic put it: "no production software will have to change as a result."
- The AES result is reduced-round only. The full cipher that protects the actual internet is untouched.
- The math still needed humans to trust it. Anthropic says it took two researchers nearly a month to gain confidence the discovered method was correct.
That last point is worth sitting with. The interesting claim here is not "AI cracks crypto." It's that a model produced novel cryptographic mathematics against a NIST candidate — a categorically different thing from finding bugs in code. Finding a missing auth check (see Part 11) is pattern-matching at scale. Discovering a previously-unexploited lattice automorphism is closer to research.
What it actually means if you run real systems
This is not a "rotate your keys" moment. Nothing about your production stack is more exposed today than it was last week.
What it is is a signal about the bar that post-quantum standardization now has to clear. The next generation of cryptographic standards — the ones being selected right now to replace what quantum computers will eventually threaten — will have to withstand AI-assisted cryptanalysis as a routine part of their review, not as an exotic edge case. HAWK just got an early, public preview of that scrutiny, and a candidate scheme lost half its margin to a model running for 60 hours.
The honest framing of this whole moment: the capability is real and genuinely new, the practical impact today is zero, and both of those facts need to travel together. Anything that separates them — in either direction — is propaganda, not analysis.
Source: Anthropic — "Discovering cryptographic weaknesses" (July 28, 2026). These are Anthropic's own published research results and, as far as we can tell, have not yet been independently reproduced; we've framed them accordingly.
Wondering what post-quantum readiness actually means for your systems? Talk to Sid Techno — we'll help you separate the real timeline from the headlines.