For about 25 years, a short set of numbers sat at the heart of everyday internet security with no clear public explanation for where they came from. Those numbers are the seeds used to generate the base points for the NIST elliptic curves – P-192, P-224, P-256, P-384 and P-521 – the curves that help protect digital signatures, secure logins and encrypted connections all over the world. On October 16, 2026, Anthropic published research saying its model Claude had found the answer, and the answer was not what most people expected.
According to Anthropic, the seeds were not hashes of a secret phrase at all. They were output from lrand48, an old random number generator from the C programming library, seeded with a Unix timestamp. The runs Claude identified date from May 1997 to April 1998. Anthropic has published reproduction code, so other researchers do not have to take the claim on faith. They can run it themselves and watch the same numbers come out.
How the question started
Elliptic curve cryptography sounds abstract until you remember how often you rely on it. Every time a browser checks a website certificate, a phone verifies a signed update, or a bank system confirms a digital signature, elliptic curves are often doing quiet work in the background. The NIST curves became standard tools for that work, which is why the origin of their seeds mattered so much.
The worry was straightforward. Curve base points are supposed to be generated in a way that leaves no hidden advantage for whoever picked them. If the seeds came from a phrase chosen in private, critics asked, could someone who knew the phrase have shaped the result? No evidence ever showed that happened, but the missing explanation left room for doubt. For a piece of public infrastructure, doubt itself was the problem.
Readers who followed Claude Mythos 5 redesigning eight quantum circuits will recognize the pattern here. Put a hard, checkable problem in front of a model, then insist on proof that can survive outside scrutiny. In the circuit work, formal proofs did the checking. Here, reproduction code does it.
A cold case, asked cold
The setup matters. Anthropic says an employee asked an internal Claude research model about the origin of the seeds, without pointing it at the old rumor. That detail is important. Claude was not fed the answer and asked to decorate it. It had to work the problem from the numbers themselves.
And it worked fast. In under half an hour, according to Anthropic, Claude found the pattern. A team of ten at Anthropic, including Nicholas Carlini, then did the human work of testing, confirming and writing up what the model had found. Speed got the lead. Careful verification turned the lead into a published result.
The old lead, it turns out, pointed the wrong way. NSA mathematician Jerry Solinas, who died in 2023, had recalled that the seeds might have come from hashing a joke passphrase, something like “Give Bob and Jerry a raise.” It was a charming story, the kind that sticks around because it sounds like something tired colleagues would type late at night. Anthropic’s research says it was wrong. The seeds did not come from a phrase at all.
The fingerprint in the numbers
So how did Claude spot it? Anthropic says the model noticed the fingerprint of lrand48 in the seed values. The generator works with 48 bits of internal state but returns only 31 of them, and its output has a tell: the top bit is zero in its 32-bit words. Once you know what to look for, that pattern stands out the way a familiar handwriting slant stands out in an anonymous note.
From there, the problem became one of recovery. Claude used LLL lattice techniques, a well established mathematical tool for finding hidden structure in numbers, to work backward from the published output to the generator’s hidden state. Once the state was recovered, the starting point could be identified as a Unix timestamp – essentially a date and time used as the seed.
The most concrete detail concerns P-256, the most widely used curve of the set. Anthropic says its seed was the 1,842nd candidate drawn from a run started on 16 August 1997. That is an oddly specific fact, and specificity is exactly what this mystery always lacked. Not a rumor. Not a recollection. A run date, a position in a sequence, and code that reproduces it.
More than one mystery in the same paper
The NIST seeds were the headline, but Anthropic says Claude kept going. The model also explained the base points of six curves from FreeLIP, a library by cryptographer Arjen Lenstra. Those numbers had their own unexplained origins, and Claude traced them as well.
Not everything gave way. Anthropic is plain about the leftovers: the B-163 seed and some base points still have no explanation. That honesty helps the rest of the paper. A result that claims to solve everything invites suspicion. A result that says “these parts remain open” sounds like researchers reporting what they actually found.
Recent coverage of Claude writing most of its own codebase focused on software production. This story is software archaeology – reconstructing late 1990s decisions from the numbers they left behind, with evidence that is fixed and public.
Why this is good news for trust
Here is the part worth holding onto. Nothing about the curves themselves changed on October 16. P-256 worked on October 15, and it works now. What changed is the quality of the public record around it. A question that invited speculation for a quarter century now has an answer backed by runnable code.
That matters for ordinary people who will never read a cryptography paper. Trust in digital signatures is a chain: standard numbers, public methods, independent checks. Replacing a shrug with a timestamp, a generator name and a reproduction script makes the chain stronger in a way anyone technical can test.
It also sets a useful example for how AI research claims should land. Anthropic attributed the work, named the team, described what the model did and what it did not finish, and published code. Compare that with the care taken in AI systems learning to forecast weather weeks ahead, where value also depends on checks other people can repeat. In both cases, the receipt matters as much as the result.
What happens next
Independent researchers will run the code. That is the point of publishing it, and scrutiny will decide how settled this answer becomes. If the seeds reproduce exactly, as Anthropic says they do, a long debate quietly closes.
There may be more cold cases like this one. Modern systems sit on layers of old decisions, some documented well, some barely at all, and a model that can work backward from a generator fingerprint could be useful far beyond one set of curves.
For now, the headline is simple and satisfying. A joke passphrase did not create the world’s most trusted curves. A dated run of an ordinary 1990s random generator did, starting one August day in 1997, with P-256 arriving 1,842 draws in. After 25 years, that is a proper answer.
Be First to Comment