Legacy cryptographic assumptions that have anchored enterprise infrastructure for decades are collapsing under the raw computational speed of frontier models, as Anthropic's Claude solved a 25-year-old cryptography mystery in under thirty minutes.
Back in 2000, NIST standardized several curves in the ECDSA standard, the most widely used framework for elliptic-curve digital signatures. The National Security Agency had supplied the seeds, the starting values used to generate the curve parameters, with zero explanation. For a quarter of a century, this lack of transparency fueled deep suspicion across the security community regarding potential backdoors, attracting prolonged amateur sleuthing efforts and a dedicated financial bounty. Even the late NSA employee Jerry Solinas had suggested the seeds were a hash of a humorous phrase, but that hypothesis ultimately proved incorrect.
Claude cracked the entire problem in under 30 minutes, identifying that the seeds originated from "lrand48," a simple C random number generator initialized with a Unix timestamp between May 1997 and April 1998. The system spotted the structural pattern because "lrand48" only outputs 31 bits, causing the top bits in some seeds to be systematically zero.
Claude cracked it in under 30 minutes.
According to Anthropic, this deterministic origin effectively renders deliberate backdoors far less likely than decades of speculation suggested. Following the resolution, Anthropic is donating the $36,864 bounty directly to the Internet Security Research Group.
This demonstration exposes how quickly automated machine intelligence can dismantle historical cryptographic obscurities that previously required years of human cryptanalysis. Engineering teams relying on legacy parameters and obscure architectural inheritances face immediate operational exposure as automated threat actors leverage identical diagnostic speeds. Organizations must audit foundational cryptographic libraries to ensure underlying dependencies withstand high-speed automated reverse engineering.