The NIST elliptic curves used with ECDSA (the Elliptic Curve Digital Signature Algorithm), published in FIPS 186-2 in 2000, each come with a 20-byte "seed" whose hash sets one of the curve's coefficients. The NSA supplied the seeds without saying where they came from, which fed two decades of suspicion that the curves might hide a deliberate weakness. The post "Claude reveals ECDSA curve origins", published on 16 October 2026 by Anthropic's Frontier Red Team, reports that an internal research model worked out how the seeds were made. The post is the version of record: it has no companion paper, and it prints scripts and outputs (released under CC0) that reproduce its claims. It is short, about 1,300 words of prose plus code. The page's metadata records a modification on 23 October 2026 without a visible note of what changed; this note follows the copy read on 13 November 2026. The authors call it "a fun closing chapter" with "no action to be taken".
The question and two terms
For P-256 the published seed is the hexadecimal string beginning C49D3608, with no stated origin. The post cites Koblitz and Menezes's survey "A Riddle Wrapped in an Enigma", an amateur investigation (a post on Steve Weis's site; Weis is one of this post's authors), and a bounty organized by Filippo Valsorda in October 2023. The best-known account, from the NSA's Jerry Solinas in 2015, was that the seeds were hashes of a humorous phrase such as "Give Bob and Jerry a raise"; he did not remember the details, the archives were not found, and he died in 2023.
A pseudorandom number generator (PRNG) produces a random-looking sequence from a starting state. lrand48, from the standard C library, is a linear congruential generator: it updates a 48-bit state by a fixed multiply-and-add and returns the top 31 bits. srand48 sets the state from an integer, filling the low 16 bits with the constant 0x330E.
How the model found the seeds
Several Anthropic employees had chased the phrase lead without success. In October 2026 an employee asked "an internal research model" how the seeds were generated, without pointing it at that clue; the post does not name the model. "In under half an hour", it reports, the model found the origin of the seeds for P-192, P-224, P-256, P-384 and P-521 ("Claude found the seeds").
The route ("How it was found"):
- Guess the software, not the phrase. The model hypothesized that the program used a weak 1990s C-library PRNG whose output can be recognized, and wrote a scanner testing whether neighboring chunks of each seed were consecutive generator outputs, over 34 combinations of generator and output slice.
- Zero bits. Written as five 32-bit words, four of the ten NIST seeds (five prime-curve, five binary-curve) had a zero top bit in all 20 words. The post gives the chance as 1/5,810, which matches the probability that at least four of ten random seeds show the pattern (our arithmetic, binomial with 1/32 per seed). The model's own reasoning, quoted in the post, says "chance 1e-6", the figure for 20 specific bits without allowing for picking four seeds of ten.
- Recover the state. A zero top bit fits
lrand48. For those four binary-curve seeds the model tried all 2¹⁷ values of the hidden bits of the first word and checked that the next four words followed. The prime-curve seeds keep only the top 8 bits of each output; recovering such a truncated generator is "a textbook lattice problem", solved with LLL basis reduction over every bit window. Five NIST prime-curve seeds and five example-curve seeds from ANSI X9.62 fit, and each recovered 48-bit state regenerates all 160 bits of its seed. - Wind back to the start. Running the generator backwards to a state of the form
srand48sets is "not evidence on its own", since such states recur every 65,536 steps. What the post treats as evidence is that all ten prime-curve starting values read as Unix timestamps between May 1997 and April 1998. The binary-curve values read as February 1984; doubled, they fall on 17 April 1998, two weeks after the P-224 run, so "the program may have halved the clock" (the code labels this "if that reading is right").
The published P-256 seed is candidate 1,842 of a run started at 1997-08-16 01:52:34 UTC (the post's script). For this note we reimplemented that one case independently from the description and obtained the published seed exactly; we did not run the post's code or check the other seeds.
The base points
Each curve also has a fixed base point, G, which the seed does not determine. The model found the origin of six: P-192 and the five ANSI X9.62 example curves of May 1997 (prime192v2, prime192v3 and prime239v1 to v3). Their x-coordinates come from the generator in FreeLIP, Arjen Lenstra's long-integer package, seeded with small integers (20000 for P-192; 20420 to 20517 for the others), discarding seven outputs and drawing candidates until one lies on the curve ("Claude found some base points too"). The post's script reports a match for all six after one to five tries. The post says "no NIST base point has been explained before", as far as the authors know; those of P-224, P-256, P-384 and P-521 do not fit.
What the authors conclude, and what stays open
The post states that this "is not a cryptographic break, new attack, or weakness in the ECDSA curves", and argues that a reproducible, clock-seeded generator makes deliberate backdoors "much less plausible" ("What this means"). Its own conclusion is stronger than that phrase: the finding "makes us more confident the curves were chosen without any backdoors or intentional weaknesses", and the authors write that "we can be confident the parameters are secure". It adds that "Most researchers trusted the curves already."
It also lists what it cannot explain ("Open questions"):
- Skipped curves. For P-256, candidate 876 of the same run also gives a suitable prime-order curve, but candidate 1,842 was published; the P-224 run skipped three such curves and the two ANSI X9.62 runs seven more. The authors do not know the criteria, "or whether this was an unintentional artifact". They judge selection of weakened curves implausible because so few were skipped, grant that curves could in principle have been chosen by searching over timestamps, and note that "Nothing in the numbers shows how many runs there were."
- The post's table scores published and skipped curves against later attack criteria. Its printed values include twist security of 2^120.2 for P-256 and 2^99.9 for skipped candidate 876, against a pass mark of 2^100; every curve in the 192-, 224- and 239-bit runs fails that criterion. The text does not comment on these values.
- B-163's seed "fits nothing Claude has tried", and the remaining NIST base points, prime and binary, are unexplained.
The bounty. The post says the $36,864 bounty will go to the Internet Security Research Group. Valsorda's announcement set it at $12,288, tripled if donated, and framed it as cracking five hashes; the post does not say how the organizer judged a finding that the seeds were not hashes.
Limitations
- Verification is computational. Each claim can be checked by running code that regenerates the published values, which the post invites; it reports no other outside review. Two authors are tied to the bounty: Steve Weis, credited in its announcement with the research that inspired it, and Amir Omidi, a contributor.
- The process is thinly documented. The model is unnamed, only a short passage of its reasoning is quoted, and the post does not say how much direction followed the first question or whether other attempts failed. In that passage the model asks whether it is "in some sort of simulated/test environment" and settles the question by its own computation ("It's verified by direct computation on my side").
- Intent is an inference. That a reproducible generator leaves little room for manipulation is the authors' argument; the skipped candidates and the unknown number of runs are the part of the record the post says it cannot explain.
Research questions
- What rule explains the skipped curves? Testing published and skipped candidates against the criteria in 1997 and 1998 drafts of ANSI X9.62 and FIPS 186-2 would show whether a known rule, or a program bug, accounts for the skips.
- How were the other base points chosen? The same search over period generators applied to P-224, P-256, P-384 and P-521 would show whether one method covers them.
- What generated the B-163 seed? A wider scan of 1990s generators and encodings, or recovered archives, would settle the one seed that resists the method.
- Is the halved-clock reading right? Other values from the binary-curve runs that fit a halved timestamp would confirm or replace it.
- How typical is this success? A record of attempts on comparable problems (other unexplained constants in standards), with prompts and failures, would show whether this case was selected.
How it connects
- Discovering cryptographic weaknesses with Claude (July 2026) reports attacks on cryptographic algorithms developed with Claude Mythos Preview and shares an author with this post (Steve Weis); here a model explains how standard parameters were produced instead. The ECDSA post does not cite it; the connection is this note's.
- The post names no model and links no system card. For the Frontier Red Team's earlier security work with a named internal model, see the Claude Mythos Preview system card; whether the model in this post is that one or a later one is not stated. The connection is this note's.
References
- Wood, J., de Valence, H., Howell, A., Voase, Z., Sweet, L., Omidi, A., Crowley, P., McGinniss, I., Carlini, N., and Weis, S. (2026, October 16). Claude reveals ECDSA curve origins. Anthropic Frontier Red Team. https://www.anthropic.com/research/ecdsa-curve-origins (code released under CC0 1.0). record
- Valsorda, F. (2023, October 5). Announcing the $12k NIST Elliptic Curves Seeds Bounty. https://words.filippo.io/seeds-bounty/. Read for the bounty's amount, terms and contributors. record
- Koblitz, N., and Menezes, A. (2015). A Riddle Wrapped in an Enigma. Cryptology ePrint Archive, Paper 2015/1018. Cited in the post; not read for this note. record
- National Institute of Standards and Technology. (2000). FIPS 186-2, Digital Signature Standard. Cited in the post; not read for this note. record