In the late 1990s the National Security Agency delivered a set of elliptic-curve parameters to the National Institute of Standards and Technology. While the curves were adopted worldwide, the source of the initial seeds—the numbers used to derive the curves—was never explained, prompting speculation that the NSA had embedded a covert weakness.
AI Investigation
An internal Anthropic research model, known as Claude, was tasked with exploring the mystery without any pre-conceptions about a hidden phrase or backdoor. Within roughly thirty minutes the model reproduced the exact sequence generated by the C library function lrand48, matching the published seed for the P-256 curve at the 1,842nd trial. The match was statistically improbable (about one in a million), leading Claude to confirm that the seeds originated from a conventional random-number routine seeded with timestamps from 1997-1998.
The result, posted by Anthropic, shows that the seeds are reproducible using publicly available code. While the researchers could not explain why the generator skipped many other viable candidates, the finding makes a deliberately inserted vulnerability far less likely.
Implications for Bitcoin
Bitcoin relies on the secp256k1 curve, which differs from the NIST curves in that it is defined by two fixed constants and leaves virtually no room for hidden parameters. Consequently, the revelation about the NIST seeds does not impact Bitcoin’s cryptographic foundation. The article notes, however, that future quantum computers could threaten any elliptic-curve system, including Bitcoin, by deriving private keys from public ones. Recent advances such as Google’s Willow chip, which dramatically reduced error rates in quantum operations, underscore the long-term relevance of that risk.
Why it matters
The resolution of a 25-year cryptographic puzzle restores confidence in the integrity of the NIST curves and demonstrates the power of modern AI to solve deep technical problems. Users of the affected curves do not need to replace keys or migrate to alternative standards, but the episode highlights the importance of transparent parameter generation. Moreover, the successful application of an AI model to cryptanalysis may shape future security audits and bounty programs, as evidenced by the $36,864 prize now earmarked for the Internet Security Research Group.
