Claude AI traces origin of 25-year NSA cryptography puzzle to lrand48

What happened

Anthropic's AI model Claude has traced the origin of a cryptography puzzle that had gone unexplained for about 25 years, according to reporting by The Decoder. The question concerns the seed values behind the standard curves that the US National Institute of Standards and Technology (NIST) specifies for digital signatures. The National Security Agency (NSA) supplied those seeds, and no rationale for them was ever given.

That silence fed long-standing suspicion that the curves might contain a deliberately built-in backdoor. Per the report, Claude concluded in under 30 minutes that the seeds most likely originated from "lrand48", a simple random number generator widely used in C programming. If that attribution holds, it offers a mundane explanation for values that had been treated as suspicious for decades, and it makes a deliberate backdoor considerably less likely.

Why it matters

The episode is a clear example of what current AI models can now do in technical investigation. A question that had resisted human scrutiny for a quarter of a century was resolved through rapid pattern-matching across code, documentation and known generator behaviour. For organisations, this signals that AI is moving beyond drafting and summarising into genuine analytical work, with implications for security review, audit and the verification of legacy systems.

It also raises a practical point for digital transformation leaders. Many enterprises and public bodies run on infrastructure whose design rationale has been lost over time. Where AI can reconstruct that rationale quickly, the cost of assessing legacy decisions falls, and the question shifts from whether a review is affordable to how it should be governed and who is accountable for acting on its findings.

By the numbers

• 25 years: the period the cryptographic question reportedly remained unresolved.
• Under 30 minutes: the time Claude reportedly took to identify "lrand48" as the likely source of the seeds.

The Renascence take

Most coverage will focus on the NSA angle, but the more useful lesson sits in the workflow. The breakthrough came from an AI system that could hold a long-running, unresolved question and reframe it in minutes, which is precisely the kind of capability that customer-obsessed teams should be testing against their own legacy knowledge gaps.

The headline is the mystery; the real story is how quickly a well-framed question can dissolve years of institutional assumption. Behavioural economics has long shown that unexplained defaults get treated as signals of intent, and this case is a textbook example of an opaque choice being read as a threat. Operators should audit their own unexplained defaults, in product settings, service rules and internal processes, and document the reasoning behind them before a capable model, or a suspicious customer, does it for them.