Anthropic’s Claude Solves NSA’s 25-Year Cryptographic Seed Puzzle in Just 30 Minutes

Thirty minutes. That’s all it took. Anthropic’s AI model Claude cracked a cryptographic puzzle that had stumped some of the world’s best security researchers for over a quarter century — tracing a set of mysterious NSA numbers back to a simple random-number program running on timestamps from 1997 and 1998.

The numbers in question were seeds: the starting values used to generate the cryptographic curves that the U.S. National Institute of Standards and Technology adopted back in 2000. NIST took those curves straight from the NSA, no questions asked — or at least, no answers given. The origins of the seeds were never disclosed. That silence was deafening to the cryptographic community. For years, researchers suspected the NSA might have hand-picked those numbers to embed a backdoor — a hidden weakness only the agency could exploit. The Snowden disclosures didn’t help. They tied the NSA directly to efforts to quietly undermine cryptographic standards, and suddenly the suspicion around those unexplained seeds looked a lot less like paranoia.

Cryptographers Neal Koblitz and Alfred Menezes dug into the issue in a 2015 paper. Security researcher Steve Weis filed public-records requests trying to get documentation. Nothing came back. A bounty of over $36,000 sat unclaimed for years, dangled in front of anyone who could explain where those seeds actually came from.

What Claude Actually Found

An Anthropic employee put Claude to work on the problem without feeding it any preconceived theories about the seeds’ origins. Within minutes — the whole process took about 30 minutes total — Claude had an answer. The seeds traced back to lrand48, an older random-number generator. Specifically, Claude matched the sequence behind the widely used P-256 curve and confirmed that its published seed was the 1,842nd candidate the generator had evaluated.

That’s a pretty specific finding. And it’s kind of a big deal, because it means anyone can now retrace those steps. The process is replicable. The transparency that was missing for 25 years is now basically restored.

But it’s not a clean, tidy ending. Questions remain. Unclear why the program skipped over other usable candidates in the sequence. Unclear how the other components of the curves got selected. Those gaps haven’t closed. What the finding does do is knock down the most alarming theory — that the NSA engineered a deliberate, exploitable weakness. The numbers came from a known, traceable generator. Not from some secret manipulation process.

See also: Binance Expands Into Stocks in Hong Kong, South Korea, Taiwan, and Brazil

Bitcoin Wasn’t Part of This

Worth saying clearly: Bitcoin uses a different curve entirely. It runs on secp256k1, not the NIST curves at the center of this controversy. Cryptographers have long favored secp256k1 partly because its design leaves minimal room for hidden vulnerabilities — the selection of its parameters is more transparent by construction. So Bitcoin holders probably don’t need to lose sleep over this particular chapter.

For everyone else using NIST curves, Anthropic’s position is that no urgent changes are needed. The seeds were public from the start. There was no secret to exploit, at least not one that’s been found.

The $36,864 bounty — yes, that’s the exact figure — goes to the Internet Security Research Group, the nonprofit behind Let’s Encrypt. Let’s Encrypt provides free security certificates to websites, and it’s hard to think of a more fitting recipient given how much of the internet’s basic security infrastructure runs through it.

Bigger Picture for Cryptographic Security

The lrand48 generator at the center of this is old. It’s the kind of random-number generator that cryptographers today would never recommend for generating sensitive keys. Predictable generators are a known risk — if an attacker can figure out the sequence, they can potentially reconstruct private keys. The fact that these curves were seeded using something like lrand48 is, at minimum, a reminder of why modern cryptographic standards demand better randomness.

More context: XRP Plummets 62.5% from 2018 Peak as Supply Surges Amid SEC Battle

And randomness is getting harder to guarantee as computing power grows. Quantum computing isn’t an immediate threat to systems like Bitcoin — current machines can’t break the keys in use today — but the direction of travel is clear enough that researchers aren’t sitting still. Cryptographic standards that look solid now might not look that way in ten or fifteen years.

For now, though, the mystery is solved. Claude found lrand48. The 1,842nd candidate matched. The bounty goes to Let’s Encrypt.

Hub: Bitcoin price, news, and analysis