Claude traces the origins of NIST cryptographic parameters
Anthropic published a reconstruction of the origins of elliptic-curve values with reproducible code. The finding addresses a historical question rather than breaking the cryptography.
Summary
A decades-old question about cryptographic parameters has received an answer that can be reproduced computationally. Anthropic reported on October 16 how an internal Claude model identified the process that generated seeds for five NIST elliptic curves: P-192, P-224, P-256, P-384 and P-521.
According to the team, the values came from a pseudorandom generator common in C libraries, initialized with dates and times. Their origin was identified in under half an hour. The publication includes code and outputs for reproduction, rather than relying solely on an account of the model’s behavior.
In practice
These curves are used in digital-signature systems. Their seeds are public values that help determine mathematical parameters; they are not users’ private keys. Reconstructing their origin does not allow someone to sign messages as another person.
The question arose because the NSA supplied the values without explaining how they had been selected. In 2023, cryptographer Filippo Valsorda announced a bounty to uncover the seeds’ origins, documenting the historical questions and the curves’ importance to web security.
Claude searched for patterns consistent with older generators, and the team published a recipe that regenerates the known values. For P-256, for example, the standardized value corresponds to candidate 1,842 in a sequence. The exact match between the reconstruction and public parameters makes the hypothesis testable.
What we still don't know
The research does not reconstruct every original decision. Some viable curves appear to have been skipped during generation, for unknown reasons. The seed of the binary curve B-163 and the origins of several base points also remain unexplained.
Anthropic interprets the result as evidence making deliberately weakened parameter selection less plausible. It does not provide historical proof of every intention involved. The company stresses that this is not a new vulnerability and does not require operational changes by users.
Why it matters
- A reproducible result allows AI contributions to be checked through independent computation.
- Establishing parameter provenance improves transparency around security standards.
- Explaining public values is different from breaking private keys or digital signatures.