Anthropic's Claude Solves 25-Year NSA Crypto Mystery

Anthropic's Claude artificial intelligence model has successfully resolved a quarter-century-old cryptographic puzzle surrounding parameters provided by the National Security Agency (NSA). Back in 2000, the National Institute of Standards and Technology (NIST) incorporated several curves into the Elliptic Curve Digital Signature Algorithm (ECDSA) standard, which remains the most widely deployed framework for elliptic-curve digital signatures today. However, the agency delivered the foundational seed values without explaining their mathematical origin, fueling decades of suspicion among researchers regarding potential hidden backdoors.

The AI system deciphered the puzzle in under 30 minutes of analysis. Claude identified that the mysterious cryptographic seeds were generated using "lrand48," a simple pseudorandom number generator written in C, seeded with a Unix timestamp ranging between May 1997 and April 1998. The model noticed the subtle structural pattern because "lrand48" outputs just 31 bits of data, causing the highest bits in certain seeds to be systematically filled with zeros. This precise insight disproved a long-standing conjecture by the late NSA employee Jerry Solinas, who had previously suggested the seeds were formed from a hash of a humorous phrase.

According to Anthropic, uncovering this mundane generation process makes the existence of deliberate mathematical backdoors in the standard significantly less likely. The breakthrough brings closure to years of amateur sleuthing and public bounty challenges aimed at determining how the values were crafted. Following the solution, Anthropic confirmed it is donating the entire $36,864 bounty award to the non-profit Internet Security Research Group.

For security practitioners and software engineers, the resolution provides welcome confidence in the structural safety of widely used ECDSA digital signatures. Beyond settling a historic debate in cryptography, the achievement illustrates the growing capability of frontier language models to audit complex legacy software artifacts, identify subtle numerical anomalies, and solve open mathematical problems that have stumped human experts for decades.