Claude solved a 25-year-old NIST cryptographic mystery in under 30 minutes

Claude solved a cryptographic puzzle that had remained open for 25 years in less than 30 minutes.

In 2000, the U.S. government published NIST elliptic curves such as P-256, which are used in ECDSA, one of the world’s most common digital signature standards. Bitcoin also uses ECDSA, but with a different curve, secp256k1, which is not part of the NIST curves.

The NSA provided these numbers without explanation, which led for years to suspicions that they had been deliberately chosen to create a backdoor. A common theory was that NSA employee Jerry Solinas generated them from a joke phrase such as “Give Bob and Jerry a raise.” He died in 2023 without recalling the details, and Anthropic staff had already tried to test that theory without success.

Then one employee asked an internal research version of Claude to figure out how the numbers were obtained, without telling the model about the suspected phrase.

Claude found the answer: there was no phrase. The numbers were generated using a common C pseudorandom number generator, with the computer’s system clock time used as the seed. The generation took place between May 1997 and April 1998. Now anyone can reproduce the numbers. More details are available in Anthropic’s research.

Anthropic emphasizes that this discovery does not break the security of existing systems and does not create new vulnerabilities. On the contrary, the theory of an intentionally embedded backdoor becomes significantly less plausible, so no changes are required.

Still, some questions remain open. For example, it is still unknown why other numbers that also worked for the task were discarded.

What is most impressive is the method itself.

Many researchers spent years looking for the supposed phrase. That hypothesis became the starting point around which almost all searches were built.

Claude took a different path. Instead of searching for a phrase, the model inferred how a program written in 1997 could generate such numbers, then tested that hypothesis.

And in doing so, it solved the problem.