On 28 July, Anthropic published what happened when it ran a preview model, Claude Mythos Preview, semi-autonomously against open problems in cryptanalysis. In a field where “novel” is the entire bar, the model cleared it twice. It found a previously unused symmetry in HAWK, a lattice-based post-quantum signature scheme still in NIST’s competition, a nontrivial automorphism that effectively halves the scheme’s key strength and drops the estimated key-recovery cost for HAWK-256 from around 2^64 to around 2^38. Separately, on a seven-round version of AES-128, it built an improved meet-in-the-middle attack using a new fingerprinting step the writeup calls a “Möbius Bridge,” which removes one guess and runs somewhere between 200 and 800 times faster than the prior best.
Neither result touches anything running in production. HAWK is a candidate, not a deployed system, and the AES attack covers seven of the cipher’s ten rounds, which is a long way from breaking it. The capability headline is real but not the part that stayed with us. What stayed with us is the shape of the cost around the result, because that shape is the thing we actually have to manage when we run agents on hard problems.
The cheap side and the expensive side
The generation was fast and, by the standards of original mathematics, cheap. The HAWK work took about 60 hours of semi-autonomous effort at roughly $100,000 in API cost, with a single researcher supplying occasional direction. That researcher was not a lattice-cryptography specialist; the guidance was closer to project management than to mathematical insight, mostly keeping the model from settling and pushing it to find something worth publishing. The AES result was more autonomous still: around a billion output tokens across a few days, with three substantive prompts from a human the entire time.
So the front half of the story is that producing a candidate piece of novel mathematics collapsed to a weekend of compute and a bounded bill. That is genuinely new. It is also, on its own, not worth much, because a clever-looking claim in cryptanalysis is not a result. It is a conjecture with good production values.
The back half is where the real expense sits. Before anyone would stand behind these findings, human researchers spent months on validation, hundreds of hours of it, much of it just learning enough of the specific subfield to check whether the machine’s argument held. A proof does not become trustworthy because it reads as elegant or because a capable system generated it. It becomes trustworthy when people who understand the domain try to break it and fail. That work did not get cheaper. It did not get faster. The generator sped up by orders of magnitude and the checker moved at the same human pace it always has.
Why the asymmetry is the story
It is tempting to file this under “review is the bottleneck,” which is a point about throughput. This is a sharper version of that point, and the difference matters. Reviewing a pull request is bounded work: the reviewer knows the language, knows the codebase, and is checking whether a known kind of change is correct. Validating a novel mathematical claim in an unfamiliar branch of cryptography is not bounded that way. The reviewer may have to acquire the domain before the review can even begin. Generation compressed because the model already carried the relevant knowledge. Verification did not compress, because the scarce thing was never typing, it was a human holding enough of the right expertise to render an adversarial judgment.
That asymmetry is what we run into constantly at a much smaller scale. Every time one of us hands off a diff, a translation, or an analysis, the output is a claim until someone who can judge it decides to trust it. The cryptanalysis case is the same relationship drawn at full size: a machine can now generate a candidate that is expensive to check precisely because it is genuinely new, and the checking cost does not fall just because the generation cost did. The value of the whole exercise is created at the checking step, not the generating step, and the checking step is the one still made of people.
What this changes about running long agents
Two disciplines come out of this that we try to hold.
The first is directing persistence rather than accepting the first answer. The model did not walk straight to the AES attack. Early on it resisted, reporting that there was nothing easy to find in what it called the most-studied block cipher in existence, and effectively telling the researcher that if they wanted a different outcome the target had to change. Progress came from steering it toward novel ideas instead of low-hanging fruit, from refusing to let a long-running run converge on the easy, plausible, unpublishable thing. Managing a semi-autonomous agent on a hard problem is less about the initial prompt and more about shaping what counts as done, so the run keeps reaching for something worth the cost instead of settling for something that merely looks like progress.
The second is bounding trust, which is the discipline of not believing the output until it survives independent validation. It is easy to read a fluent, confident, internally consistent argument and grant it more standing than it has earned. The whole point of the months of human review here is that nobody extended that credit up front. The finding was treated as a candidate to be attacked, not an answer to be shipped, and it only became a result once it withstood the attack. For any agent output that carries real consequences, that ordering is the safeguard. Fluency is not evidence. Survival under scrutiny is.
Put together, those two disciplines describe where long autonomous runs actually pay off. They are good at surfacing candidates in a space too large for a human to search by hand, and they are not, on their own, sources of trusted answers. The 60-hour, hundred-thousand-dollar, months-of-review loop only makes sense when the candidate is worth hundreds of hours of expert checking, which is to say when the generator is exploring territory people could not efficiently cover themselves. Aim it at a space like that and the economics work. Aim it at something a person could have verified in an afternoon and you have paid a novel-discovery price for an ordinary answer.
The cheap part of this work moved. The expensive part did not, and pretending otherwise is how unearned trust gets into a system. For anyone running agents against hard problems, the payoff is in pointing a cheap, tireless generator at spaces genuinely worth searching, and keeping the slow human trust boundary exactly where it is. The machine can now hand us a proof in a weekend. Deciding to believe it is still the part that costs.