Anthropic says its internal AI model Claude solved a 25-year-old mystery about cryptographic seeds used in ECDSA standards in under 30 minutes

Anthropic announced on October 16 that its internal AI research model Claude resolved a long-standing question about how certain numerical values, known as seeds, were generated for use in cryptographic standards related to the Elliptic Curve Digital Signature Algorithm (ECDSA). The seeds were originally provided by the U.S. National Security Agency (NSA) and are included in standards published by the National Institute of Standards and Technology (NIST) in 2000. For over two decades, the method used to generate these seeds remained unknown, raising concerns that the NSA may have intentionally selected curves with hidden weaknesses. A former NSA employee recalled in 2015 that the seeds were derived from a hashed English phrase, but could not remember the exact process, and no documentation was available to confirm the claim.

According to Anthropic, a team member asked Claude to investigate the origin of the seeds, and the model identified the technique within 30 minutes. The seeds were generated using a pseudorandom number generator called lrand48, found in a C language library, which used timestamps from 1997 and 1998 as initial values. By combining these values, Claude was able to reproduce the original seeds. Anthropic also released the code used for reproduction. The discovery applies to five curves listed in one NIST standard and four additional curves adopted by the agency. In some cases, Claude also determined how basis points used in public key calculations were selected. However, the finding does not indicate that any cryptographic systems have been compromised. The seeds are public values defined in the standards and are distinct from users' private keys.