Skip to content

Anthropic

Claude traced NIST curve seeds to 1990s timestamps

6 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: anthropic.com

On October 16, Anthropic said that Claude had traced the unexplained NSA seeds behind NIST’s P-192 through P-521 curves to the C library’s lrand48 random-number generator, started from a Unix timestamp. For code that signs with ECDSA on these curves, nothing changes: Anthropic says the finding is not a break, a new attack or a weakness, there is nothing to do, and our own rerun reproduces every prime-curve seed down to the second. The B-163 seed and the base points of P-224, P-256, P-384 and P-521 are still unexplained.

The seeds come from FIPS 186-2, the Digital Signature Standard that NIST adopted in 2000. Each curve’s coefficient b is the hash of a published 20-byte seed. The NSA supplied those seeds with no explanation (P-256’s is C49D3608 86E70493 6A6678E1 139D26B7 819F7E90). In October an Anthropic employee asked an internal research model how the seeds were generated, without pointing it at any clue. Claude worked it out in under half an hour.

A passphrase, or the clock?

Until now everyone had the same story. On October 12, 2023, Schneier on Security passed on the rumor that Jerry Solinas at the NSA had made the seeds by running some variation of “Give Jerry a raise” through SHA1. The point of hashing was to wipe out any structure in the seed, so the NSA could not have chosen a deliberately weak one. The post called recovering the seeds a potentially devastating blow to the theory that the curves were generated maliciously. It also noted that these are not the constants of Dual_EC_PRNG, the generator the NSA backdoored. Solinas had told the same story in 2015: “The message was along the lines of ‘Give Bob and Jerry a raise’ or ‘Bob and Jerry rule’ or something like that.” He died in 2023.

That story was wrong. None of the seeds is a hash. Several Anthropic employees had gone after the phrase and found nothing, and Steve Weis, one of the authors, wrote on Hacker News that he had confirmed the passphrase account with four people. Koblitz and Menezes covered the wider controversy in “A Riddle Wrapped in an Enigma”.

How Claude got there

Claude dropped the phrase hunt. Instead it guessed that the software had used a weak generator from a 1990s C library, the kind whose output can be recognized. It wrote a scanner covering 34 combinations of generator and output slicing. The tell was in the top bits: written as five 32-bit words each, four of the ten NIST seeds had the top bit clear in every word. That is the signature of lrand48, which returns only the top 31 bits of its 48-bit state. For those four binary-curve seeds, Claude brute-forced the 17 hidden bits. The prime-curve seeds kept fewer bits of each output (lrand48() >> 23), so recovering the generator state became a textbook lattice problem, solved with LLL basis reduction.

Claude then ran each state backwards to the form srand48 sets, with the low 16 bits equal to 0x330E. Ten prime-curve seeds (five NIST, five from ANSI X9.62) all came out as Unix timestamps between May 1997 and April 1998. The binary seeds read as February 1984. Doubled, they land on April 17, 1998, two weeks after the P-224 run, so the program may have halved the clock. Before trusting the result, Claude asked itself: “Let me be skeptical: am I in some sort of simulated/test environment?”

Some base points also fell out. The base point G is the fixed point that every ECDSA public key is a multiple of. For P-192 and five ANSI example curves, all generated in May 1997, G comes from the random generator in Arjen Lenstra’s FreeLIP package, seeded with small integers (20000 for P-192). Anthropic says no NIST base point had been explained before.

Skipped curves and B-163

Some workable candidates were passed over. P-256 is candidate 1842 of its run, yet candidate 876 also gives a prime-order curve of the right size. The P-224 run skipped three such curves and the ANSI runs seven more, by criteria Anthropic does not know. Its argument against a planted weakness rests only on how few curves were skipped: if weak curves were that easy to hit, weak curves would have to be common. It concedes that searching over timestamps for some property is conceivable in principle, and that nothing in the numbers shows how many runs there were. B-163 fits nothing Claude tried. The base points of P-224 through P-521 and of the binary curves are still unexplained.

Our analysis

What we checked

On October 18 we ran Anthropic’s seed script, rebuilt the base-point recipe from FreeLIP’s own source code, and recomputed the post’s odds and its figures on skipped curves.

How we check stories

The core claim holds: each of the ten prime-curve seeds comes back byte for byte from its stated UTC timestamp, and a one-second shift or a neighboring candidate breaks the match. Two numbers need narrowing. The 1 in 5,810 is the chance that at least four of ten seeds show clear top bits; Claude’s own “20 bits, 1e-6” applies to four specific seeds. And one of the seven skipped ANSI candidates fails a test that was public before 1997, which leaves six unexplained.

  • We recovered B-233, B-283, B-409 and B-571 from the seeds alone, at candidates 155, 32, 60 and 73. Because the clock was halved, each timestamp is ambiguous by one second. B-163 has the top bit set in words 1, 3 and 5, so it cannot be raw lrand48 output.
  • All six May 1997 base points match OpenSSL’s published values. Discarding seven outputs is the only count from 0 to 10 that works for P-192, an empirical fit.
  • Schneier’s $12K was the cash prize of $12,288. The $36,864 is that amount tripled for a donation, per Filippo Valsorda’s bounty page, and it goes to the Internet Security Research Group.
  • The failing ANSI candidate, 239-bit #1,562, fails the MOV transfer check, which was published in 1991-94. The skipped P-256 candidate #876 has twist security of 2^99.9 against a 2^100 bar, but that attack dates to August 2000, after the run.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.