Summary: Anthropic’s Frontier Red Team reported on July 28 that Claude Mythos Preview advanced cryptanalysis of HAWK and a seven-round reduction of AES-128. It reduced the expected HAWK-256 key-recovery cost from 2^64 to 2^38 after 60 hours and created a technique speeding reduced AES attacks by 200 to 800 times. Neither result breaks full AES-128 or affects deployed systems, but each attack cost about $100,000 in API fees.
On July 28, Anthropic’s Frontier Red Team released “Discovering cryptographic weaknesses with Claude,” a report describing what happened when Claude Mythos Preview was pointed at attack research on two well-known cryptographic targets: HAWK, a post-quantum signature scheme, and a round-reduced form of AES-128. Both produced real progress, which makes this one of the most concrete demonstrations to date of automated vulnerability research with AI. The framing deserves as much weight as the numbers: here was not a model answering cryptography exam questions, but a model grinding on an open research problem for days and returning a result the field has to take seriously.
HAWK: 60 Hours, 2^64 Down to 2^38
HAWK belongs to NIST’s additional signatures process — the track that will standardize post-quantum signature schemes past the first round of selections. It comes out of the Falcon line of work and chases signatures small enough to run on constrained devices, which is precisely the sort of design whose implementation subtleties and analysis margins draw years of scrutiny before standardization.
As the report tells it, Claude Mythos Preview moved past the best-known key recovery attack against HAWK after 60 hours of work in total: the expected attack cost against HAWK-256 fell from 2^64 to 2^38 — 26 bits gone. Read on a logarithmic scale, that separates a computation nobody could run from one that is merely enormous. A record human cryptographers assembled across years of public analysis moved in under three days of model work time.
AES: Fingerprinting by Way of Möbius Bridge
The second result sits on the symmetric side. Mythos built a fingerprinting technique it calls “Möbius Bridge,” which speeds up attacks on 7-round reduced AES-128 by a factor of 200 to 800. Fingerprinting is an old move in symmetric cryptanalysis: tag intermediate states so partial information can be matched and reused across rounds or encryptions, squeezing signal out of what otherwise reads as noise. The novelty is authorship — the technique emerged not from a human researcher’s notebook but from model-generated analysis that humans then verified. One clarification: the variant under study has its rounds deliberately cut to seven. Full 10-round AES-128 was not broken, and everyday encrypted traffic is untouched.
One Billion Tokens and a $100,000 Bill
The mechanics deserve recording too. Each of the two main attacks ran up roughly $100,000 in API fees. The AES discovery emerged only after Claude put out about one billion output tokens across several days of largely autonomous work. Call the workflow “humans set the goal, the model runs the experiments”: human cryptographers marked out the attack surface and the evaluation criteria, while the model carried the volume of trying, verifying, and iterating. That bill lands below what a comparable human team would burn, yet it is nowhere near pocket change — and the price point is itself information. Attack research at this level now runs on a meter, denominated in tokens instead of headcount, and anyone with a budget in the low six figures can commission a serious attempt at an open cryptanalysis problem.
Why Deployed Systems Are Still Safe
Neither result affects deployed systems, the report says plainly. HAWK is still in evaluation at NIST and has not seen wide adoption; the AES attack reaches only the 7-round variant. Standards bodies work from the assumption that attacks only get better over time, and this report is a reminder of how fast “better” now arrives — a candidate scheme’s safety margin got repriced in a summer, before it ever shipped.
The real signal sits on the capability curve. Hand a model a clear goal and the right tools, and it can push research forward in cryptography — a field with almost no tolerance for error, where a wrong claim wastes a reviewer’s month and a right one reshapes a standard. Anthropic and the UK’s AISI have both previously assessed Mythos Preview’s cybersecurity capabilities; this new report lays the offensive results on the table. For defenders, three things follow. Attack research scales with compute budgets rather than researcher headcount. Safety margins on reduced-round and niche schemes should be treated as shorter than the literature suggests. And red teams using AI to find weaknesses is no longer a thought experiment — the defensive side of the house will have to run the same kind of autonomous analysis just to keep pace.
Sources
- Anthropic — Discovering cryptographic weaknesses with Claude
- red.anthropic.com — Assessing Claude Mythos Preview’s cybersecurity capabilities