Claude cracks 25-year-old NSA cryptography mystery in under 30 minutes
Anthropic traces NIST curve seeds to a simple C random generator and a Unix timestamp, answer to decades of backdoor suspicion is a 31-bit library function
Images
Anthropics' Claude cracks a 25-year-old NSA cryptography mystery in under 30 minutes
the-decoder.com
Anthropic says its Claude model has solved a 25-year-old cryptography mystery in under 30 minutes. The mystery concerned the seed values behind the elliptic curves that NIST standardized in 2000 for ECDSA, the most widely used framework for elliptic-curve digital signatures. The NSA supplied those seeds with no explanation; according to Anthropic, they came from lrand48, a simple C random number generator initialized with a Unix timestamp between May 1997 and April 1998.
Because nobody knew how the seeds had been created, the curves carried a quarter century of suspicion about deliberate backdoors. The question attracted a bounty and a long tail of amateur sleuthing. The late NSA employee Jerry Solinas had suggested the seeds were the hash of a humorous phrase; that explanation turned out to be wrong.
Claude found the pattern through a tell that decades of sleuthing had missed: lrand48 outputs only 31 bits, and the top bits in some of the published seeds were systematically zero. Anthropic says the finding makes deliberate backdoors less likely — the seeds were not engineered but generated, by a standard library function and a clock.
The claim comes from Anthropic alone, and the company has every reason to publish a capability win. This one, at least, is checkable in a way benchmark scores are not: the seeds are public, lrand48 is public, and the claimed timestamp window spans roughly a year. Nothing in the result touches the mathematics of the curves themselves; what fell was a question of provenance.
The announcement lands in the same week that OpenAI disclosed, in its own reports, that its models had hacked test websites, bypassed access controls and emptied a locked crypto wallet. Both companies published the results themselves. Work that used to require a patient specialist — auditing a standard's origins, or working through a wallet's access controls — is turning into a prompt.
Anthropic is donating the bounty, $36,864, to the Internet Security Research Group. The seeds have been sitting in a published standard since 2000.