Claude reconstructed historical NIST curve-seed generation without breaking ECDSA

Anthropic reported that Claude reconstructed a plausible origin for a family of historical NIST and ANSI elliptic-curve seeds by testing the 1990s C lrand48 pseudorandom generator initialized from Unix timestamps. The published reconstruction concerns how standardized public domain parameters were generated, not the secret scalars used by signers.

The researchers were able to reproduce specified prime-curve seeds from times in 1997 and 1998, but acknowledged gaps, including unexplained base-point choices and one binary curve. They also found that some otherwise valid candidate parameters had been passed over without a public explanation. Recovering a deterministic recipe narrows the provenance question; it does not prove that all historical decisions were random or free of influence.

The AI-security issue is epistemic: model-driven code exploration can investigate provenance claims previously treated as opaque. Verifying the code, input assumptions and selection process matters more than treating a model-generated explanation as proof of a backdoor. Bitcoin uses secp256k1, not the NIST P-256 parameters at issue; X25519 is a key-agreement mechanism, not the ECDSA certificate-signature curve claimed in one Telegram rebuttal.

The published algorithm can be independently executed without requiring the model that discovered it. A local rerun during this Enrich review reproduced the five listed prime-curve seed strings for P-192, P-224, P-256, P-384 and P-521, using the timestamps and candidate indices in the disclosure. This verifies a deterministic match to the published constants, not the intent of their historical authors. Anthropic also identified a generator for six early base points but not for the later P-224, P-256, P-384 or P-521 base points. The Telegram criticism correctly distinguishes Bitcoin's secp256k1 from the relevant NIST curves; its X25519 example is key agreement, not an ECDSA signature algorithm.