Reconstructing a parameter history
Anthropic reports that an internal research model helped recover how several widely used elliptic-curve seeds were generated. The investigation concerns the historical origins of parameters used in ECDSA curves, including NIST P-192, P-224, P-256, P-384 and P-521.
According to the company, the recovered seeds came from the C library's lrand48 random-number generator initialized with Unix timestamps. The timestamps point to a period from May 1997 to April 1998. Reconstructing that process helps explain numbers whose generation history had remained uncertain.
What Claude helped recover
The report describes analyzing a linear congruential generator and using lattice-based reasoning to reconstruct its state. The team then checked whether the resulting output reproduced the published seeds. Anthropic says the model recovered the relevant generation pattern in under half an hour.
The broader investigation also examined base points and other curves. Its account is not a complete explanation of every parameter: the B-163 seed remained unexplained, as did the origins of several base points and some choices about which candidate values were skipped. Those open questions are part of the result.
This is an origin finding
Anthropic explicitly says the work is not a cryptographic break, a new attack or a weakness in the ECDSA curves. Recovering the generator behind a public parameter is different from recovering a user's private signing key. The report should not be read as an announcement that Claude can defeat ECDSA signatures.
Franklin's reading is that the interesting AI capability here is reconstructing a constrained historical process and checking it against exact published values. Readers evaluating the work should separate the parameters successfully reproduced from those still unexplained. The source's reproducible reconstruction is more informative than a broad claim that AI has broken encryption.