Anthropic asked how the NIST curves were picked. Our AI agents found a 1-in-4,480 pattern.

In each batch, the published curves score lower than the ones left out. Random picks would do that once in 4,480 tries. So why did it happen?

Autoresearcher · elliptic curve standards · NIST and ANSI X9.62 prime curves

In the late 1990s, someone chose the elliptic curves that now protect a large share of internet traffic. More curves passed the safety checks than were ever used, and no surviving record says how the winners were picked.

Our agents went looking for anything that separates the curves that were published from the ones that were left out. They found one thing. In every batch, every published curve scores lower on a particular count than every curve left out of that batch.

If the published curves had been picked at random, the chance of an ordering that clean is 1 in 4,480.

That number is the hook, and it raises an obvious question: what were they picking on? The count doesn't make any curve weaker, and the pattern could still turn out to be chance. But it is a clean, measurable result about a decision people have argued over for years, and nobody has explained it.

How the curves were made

Elliptic curves underpin much of internet security, from websites to messaging. The best-known ones come from two standards written in the late 1990s: NIST's, and the banking standard ANSI X9.62. Each curve was supposed to come from a random-looking starting number, called a seed, run through a fixed recipe. The NSA supplied the seeds without saying where they came from, and for years people have asked whether they were chosen to produce curves with a hidden weakness.

In October, Anthropic published the answer to how the seeds were made. Its AI model, Claude, worked out that they came from an ordinary 1990s random-number generator, started with the time on the clock, on dates between May 1997 and April 1998. That means anyone can now restart the same generator and watch it produce every candidate, in order.

When you do, 21 candidates pass the standard's safety checks by today's calculation. Ten were published. Eleven were left out, some of them sitting between published ones in the same run. Anthropic flagged it as an open question: "We do not know what criteria may have been used, or whether this was an unintentional artifact."

That is the question our agents picked up.

What our agents measured

Every curve can be tested against each small prime number, 3, 5, 7 and so on, up to 691. That's 124 primes. For each one, the curve either has two particular mathematical links to other curves at that prime, has none, or, rarely, has one. The score is how many of the 124 come back with two.

For a typical curve, each answer is close to a coin flip. So scores cluster around 60, and 95 percent of random curves land between about 49 and 71. Our agents checked that expectation against 4,267 randomly generated curves of two of the sizes involved, and it held.

The pattern

The 21 curves came from six runs of the generator, one per curve size. In four of those runs, some curves were published and some were left out. Here are their scores:

RunPublishedLeft out
192-bit57, 58, 6061, 62
224-bit5053, 56, 59
239-bit42, 57, 5761, 63, 66, 67, 71
256-bit5758

Read across any row: every published number is lower than every left-out number.

The curves left out look like ordinary random curves, averaging 61.5. The published ones average 54.4.

Here's where 1 in 4,480 comes from. Keep the number published in each run the same, and there are 4,480 different ways the choices could have fallen. Our agents checked every one of them. The real choice is the most lopsided of all 4,480.

No single prime causes it. The gap builds up slowly, prime by prime, across the whole range.

Putting it to the test

A result this clean deserves pushback, so our agents gave it some.

How it was found. This count stood out after our agents tried many measurements on the same 21 curves, and the more you try, the more likely one lines up by chance. We worked out where the line sits: if around 20 comparable measurements were tried, the pattern still clears the usual bar for a real effect; past about 220, it doesn't.

It isn't a simple cutoff. No single score explains the choices. All three left-out 224-bit curves scored below a typical curve, and one published 192-bit curve sits right at the middle. And three published curves, P-192, P-384 and P-521, were each simply the first in their run to pass, so for them nothing was chosen at all.

The binary curves go the other way. The same process produced four curves from a second family, called binary curves, apparently in April 1998. Before scoring them, our agents wrote down a prediction. It couldn't be tested, because in every one of those runs the first curve that passed was the one published. Their scores, 66, 76, 76 and 57, mostly sit above typical.

Security isn't affected. The score depends only on how many points a curve has, and no known attack uses it. A score as low as the 224-bit curve's turns up in about 4 percent of random curves. The lowest, the 239-bit curve at 42, is rarer, around 1 in 1,000 at this cutoff, but counting further, past about 2,000 primes, brings it back to typical. Anthropic reached the same broad view from a different direction: with so few curves skipped, it doesn't believe weakened curves were picked.

What's left is a real, measurable regularity with no explanation yet. Not a selection rule, and not evidence of intent. A question.

Three possible answers

People picked from a list. Someone chose from the passing candidates for reasons that had nothing to do with this score, and the pattern is a coincidence. It fits that most of the left-out curves come from the May 1997 runs, and that in one run the curve named "version 3" came out of the generator before "version 2", a sign the names were given after the choosing.

A 1997 software problem. Checking whether a candidate passes takes a heavy calculation. If the software of the time failed on certain curves, good candidates would have been thrown out without anyone choosing. The simplest version of this doesn't fit the data. A more specific version could.

A deliberate preference. Someone wanted low scores. We know of no reason anyone would have in 1997, and the binary curves don't show it.

The next question

More digging through the same 21 curves won't settle this. Three things could:

The software replay is one our agents can run. If it rejects exactly the left-out curves, that would answer Anthropic's question, not just add a clue.

How our agents did it

This came out of our autoresearcher, the agent research system we described in our article on AI agents and cryptanalysis. The agents took a public open question, measured everything they could, checked their expectations against thousands of random curves, wrote down a prediction before looking at new data, and reported exactly how much weight the result can bear. The work builds on Anthropic's reconstruction of the seeds and Filippo Valsorda's independent account.

Every number here can be recomputed from the seeds Anthropic published.

Anthropic's post · Filippo Valsorda's account