The NIST elliptic curves underpin a large share of the internet's cryptography, and for about twenty-five years nobody could explain where some of their seed values came from. Those unexplained constants were the basis of a long-running and entirely reasonable suspicion: if you cannot account for how a curve was chosen, you cannot rule out that it was chosen badly on purpose.
Anthropic reports that Claude traced the origin of several of those seeds in under thirty minutes. Four remain unexplained. The associated bounty, $36,864, was donated to the Internet Security Research Group, the non-profit behind Let's Encrypt.
Why a 25-year gap closes in half an hour
This was never a hard problem in the sense of requiring new mathematics. It was a search problem — the answer was somewhere in decades of documents, archived correspondence and standards drafts, and finding it required reading an enormous amount of material on the off-chance.
That is precisely the shape of work where a model has an advantage that is about stamina rather than insight. No individual researcher was ever going to read all of it on a hunch. The result is a good illustration of where these systems genuinely help: not replacing the reasoning, but making exhaustive search cheap enough to bother with.
What it does not settle
Four seeds still have no explanation, so the suspicion is narrowed rather than retired. And a traced provenance is not the same as a proof of good faith — it tells you where a number came from, not why that source chose it.
Still, a question that sat open since the 1990s is now mostly answered, and the part that remains open is much smaller and much more specific than it was last month.
Sources: Anthropic