Claude traced the NIST curve seeds: aim AI at your legacy code

A 26-year cryptography mystery fell in under half an hour. On October 16, Anthropic's Frontier Red Team published how Claude traced the NIST elliptic curve seeds, the unexplained numbers behind P-256 and its siblings, to an ordinary C random number generator seeded with a Unix timestamp. Your TLS is fine. The bigger story is what AI code analysis can now do with old, undocumented systems, including yours.

What actually happened

Per Anthropic's write-up:

• The mystery. Since 2000, nobody could explain where the seeds for the NIST prime curves came from. Rumors pointed to a hashed passphrase. An earlier group of Anthropic employees chased that rumor and got nowhere.
• The finding. Claude concluded the P-192, P-224, P-256, P-384 and P-521 seeds are not hash outputs. They are output from the C library's lrand48 generator, seeded from the system clock. It found this "in under half an hour."
• How. It spotted a statistical pattern in the seed bits (Anthropic puts the odds of chance at 1 in 5,810), matched it to a generator, and rebuilt the process.
• The limits. The binary-curve seeds have a likely but unconfirmed explanation. The B-163 seed and the base points for most curves are still unexplained. Anthropic published CC0 reproduction code so anyone can check it.
• The money. The $36,864 bounty goes to the Internet Security Research Group, which runs Let's Encrypt.

Anthropic is direct about what this is not: "not a cryptographic break, new attack, or weakness" in the curves. Its advice for readers: no action needed. If anything, a clock-seeded generator makes a hidden backdoor less plausible.

Why it matters for your business

Do not rotate keys over this. Do pay attention to the method.

Every business we have worked with has its own version of the NIST seeds: a pricing rule nobody can explain, a cron job from 2014, a magic number in the shipping calculation, a plugin the last developer patched by hand. People leave that stuff alone because figuring it out costs a week of senior time.

That cost just dropped. A model that can reverse-engineer a 1997 random number generator from its output can read your old PHP and tell you what the discount logic actually does.

How we run that kind of audit:

• Ask for evidence, not a story. Make the model show the line, the input and the output that prove its claim.
• Reproduce it. Anthropic shipped code anyone can run. Hold your AI findings to the same bar: a test that passes, or it did not happen.
• Start with what costs you money. Pricing, tax, shipping and inventory logic first. Comments and naming last.
• Keep secrets out of the prompt. Strip keys and customer data before any code leaves your machine.

Sources: Anthropic Frontier Red Team.