Claude Solves 25-Year-Old NSA Crypto Puzzle in Under 30 Minutes

Anthropic's AI model Claude has solved a 25-year-old cryptography puzzle in under 30 minutes: the origin of the unexplained seeds behind NIST's standard curves for digital signatures, values supplied by the NSA decades ago without explanation.

Claude determined that the seeds came from "lrand48," a simple C random number generator. That finding, arrived at in less than half an hour, makes deliberate backdoors in the curve seeds far less likely, according to the report.

The stakes are considerable. Digital signature standards built on NIST's curves protect communications and transactions across the internet, and the NSA's unexplained seeds have fed suspicion of built-in backdoors for a quarter century. A tool that resolves that question in minutes demonstrates a new capability class for AI-assisted cryptanalysis.

What was the mystery?

NIST's standard curves for digital signatures rely on constants known as seeds. The NSA provided those seeds when the curves were standardized, but never explained how it had generated them. That omission mattered.

Unexplained constants in cryptography invite a specific fear: that the creator chose them in a way only the creator can exploit. If an adversary selected a seed through a hidden process, the adversary might retain knowledge that weakens the curve for everyone else while leaving no public trace. Cryptographers could verify that the published curves matched the published seeds, but they could not verify how the seeds themselves were born.

The suspicion persisted for 25 years because the question resisted ordinary analysis. The seeds were simply numbers. Absent an explanation from the NSA, researchers could only test candidate generation methods one by one and see whether any produced the published values.

That is the gap Claude closed. The model identified "lrand48" as the source of the seeds, working from the puzzle to the answer in under 30 minutes.

Why does the answer matter?

The finding cuts against the backdoor hypothesis. "lrand48" is a simple, well-known random number generator from the C standard library. It is not a mechanism engineered to hide secret knowledge.

If the NSA had generated the seeds through a process it alone controlled — one designed so that only the agency could reproduce or exploit the values — the discovery of a backdoor would confirm the worst assumptions about the curves. Instead, Claude traced the seeds to a mundane generator, a result that makes deliberate backdoors far less likely.

The answer does not perform a security upgrade. Nothing changes in the curves themselves. What changes is the accounting: a 25-year-old question about the provenance of widely deployed cryptographic constants now has a concrete, checkable answer.

For the engineers who build on NIST's curves, that is not a small thing. Trust in cryptographic standards rests on the absence of unexplained choices. Each unresolved constant is a standing liability, and this one has now been retired.

What does this say about AI and cryptography?

The result carries weight beyond a single puzzle. A question that stayed open for 25 years fell to an AI model in under 30 minutes.

The comparison is the story. Generations of cryptographers knew the question, knew the published values, and could not close the gap. Claude, given the same problem, searched the space of possible generation methods and landed on "lrand48" quickly enough that the solve registers as a demonstration of capability rather than a lucky break.

The finding lands at a moment when AI models are moving from code generation and text analysis toward substantive technical discovery. Cryptographic constants are a natural test bed for that shift: the problems are precisely stated, the answers are exactly verifiable, and the search space is large enough that raw persistence at scale pays off.

Anthropic has positioned Claude as a model capable of extended reasoning on technical tasks. A 30-minute solve of a decades-old NSA puzzle is the kind of concrete, dated result that will pressure other AI labs to test their models against similar open problems.

Who is affected?

The immediate audience is the cryptographic research community, which has carried the seed question since the curves were standardized. The answer gives researchers a documented resolution where before there was only an open question and an unresponsive agency.

Standards bodies face the sharper lesson. The NSA supplied the seeds without explanation, and that single act of opacity generated 25 years of suspicion against curves that may have been sound all along. The episode argues for generation methods that are transparent at the moment of standardization, so that no future constant requires a quarter century and an AI model to explain.

Users of the NIST curves — which is to say, most users of digital signatures — gain no new protection and face no new risk. The curves perform as they did before. What they gain is cleaner provenance for the values underneath.

What comes next?

The report's finding is that Claude traced the seeds to "lrand48" and did so in under 30 minutes, reducing the likelihood of deliberate backdoors. That result now invites verification by the cryptographic community, and it establishes AI models as instruments capable of closing questions that human analysis left open — a precedent other labs and researchers will test against the remaining unexplained constants in the standards record.

By Rebecca Stone, AI In Context