What was done
Anthropic researchers had Claude work out how the NIST and ANSI elliptic curve seeds were generated. It matched them to lrand48 output seeded by timestamps, and the page includes reproduction code.
Anthropic's Frontier Red Team reports that Claude explained how many of the published NIST and ANSI elliptic curve seeds were made. The seeds are not hashes of chosen phrases. They are output of the 1990s lrand48 generator, started from Unix timestamps in 1997 and 1998. The authors say this is not a cryptographic break.
How it was done
Claude scanned the published seeds for the output of period random number generators and noticed zeroed top bits, a sign of lrand48. It recovered generator states by brute force for some seeds and by lattice reduction for truncated outputs, then stepped them back to the srand48 initial values. For the binary curves the timestamps appear halved. It also traced base points of six curves to the FreeLIP random number generator. The page names an internal research model and says the first seed finding took under half an hour.
Results
The seeds of the prime curves and of B-233, B-283, B-409 and B-571 were explained, along with base points for P-192 and some X9.62 example curves. The B-163 seed and the base points of P-224, P-256, P-384 and P-521 remain unexplained. Two Python scripts that reproduce the results are published on the page under CC0.
