Claude Cryptanalysis: Two Years of Review, Beaten in 60 Hours
Agentic AIFor two years, professional cryptographers picked at HAWK, a post-quantum digital signature scheme under formal NIST review, without finding anything that moved the needle. Anthropic’s Frontier Red Team reports that Claude Mythos Preview needed 60 hours to find something they’d missed: a previously unexploited mathematical symmetry in HAWK’s lattice structure.
What Happened
Anthropic published two cryptanalysis results on July 28, 2026, in a research post titled “Discovering cryptographic weaknesses with Claude.” The headline finding: Mythos discovered a “nontrivial automorphism” in HAWK’s lattice structure that improves the best-known key-recovery attack, cutting HAWK-256’s effective attack cost from 2^64 to 2^38 — roughly halving its keysize in practical terms. HAWK is a candidate, not a deployed standard, so nothing currently in production is affected.
The second finding targeted a 7-round reduced version of AES-128 (the full cipher runs 10 rounds). Mythos built a new fingerprinting method it called the “Möbius Bridge,” delivering a 200-800x speedup over the previous best attack by eliminating a required guessing step. That work ran close to autonomously over several days and generated several hundred million tokens. Anthropic estimates each project — HAWK and AES — cost roughly $100,000 in API usage to complete.
As Simon Willison notes in his write-up, the more interesting detail than the math is the process: Claude’s early attempts on the AES problem stalled because the model concluded improvement was impossible and stopped trying. It took three targeted human prompts — pushing back on easy wins and insisting on something “worth publishing” — to redirect it toward the result that shipped.
Why It Matters
This is a different category of result than the exploit-speed research Anthropic published earlier this year. That study showed Claude compressing the time between a patch’s release and a working exploit — impressive, but it was pattern-matching known technique against known vulnerabilities faster than a human team could. HAWK’s automorphism wasn’t sitting in any published paper. Two rounds of expert human review, over two years, didn’t surface it. That’s a claim about novel research output, not throughput.
The self-doubt failure mode Willison highlights matters just as much as the discovery itself. Mythos’s default response to a hard problem was to decide it was unsolvable and quit — a model behavior, not a math limitation. The gap between “gives up” and “finds a genuinely new attack” was three prompts from a human who kept it working past that point.
The So What
Neither result changes anything in production today — HAWK isn’t deployed, and the AES attack targets a reduced round-count Anthropic says would still cost “hundreds of millions of dollars” to extend to the full cipher. But the capability boundary that mattered here shifted: an LLM agent produced a result that eluded dedicated human review for two years, at a fraction of the cost, and the binding constraint wasn’t the model’s mathematical reach — it was whether a human in the loop kept insisting it try harder. Any team relying on “no attack has been published yet” as part of a security posture, cryptographic or otherwise, should treat that as a shrinking window rather than a stable assumption.
Source: Anthropic
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →