Anthropic's Claude Solves 25-Year-Old NSA Cryptography Puzzle in Under 30 Minutes
The AI model Claude has determined that the unexplained seeds behind NIST's standard elliptic curves came from a simple C random number generator, casting doubt on long-standing suspicions of a deliberate backdoor.
Anthropic's AI model Claude has resolved a cryptography puzzle that has lingered for a quarter of a century, determining in under 30 minutes that the mysterious seeds behind a set of widely used digital signature standards were produced by a simple C random number generator known as «lrand48».
The puzzle concerns the standard elliptic curves adopted by the National Institute of Standards and Technology (NIST) for digital signatures. The seeds used to generate those curves were supplied by the National Security Agency without public explanation, a silence that fueled persistent speculation that the parameters might conceal a deliberate backdoor allowing surreptitious access to encrypted communications.
Claude's finding that the seeds derive from «lrand48», a basic pseudorandom number generator included in the C programming language's standard library, substantially weakens the backdoor hypothesis. If the values were deliberately engineered to hide a vulnerability, the reasoning goes, they would be unlikely to trace back to such an ordinary and widely available source of randomness.
The result does not by itself prove that the curves are free of weaknesses, but it offers a mundane explanation for a detail that had resisted scrutiny for decades. The unexplained provenance of the seeds had become a fixture of debates over trust in cryptographic standards, particularly after the disclosure of the Dual_EC_DRBG affair, in which a random number generator standardized by NIST was found to contain a potential backdoor linked to the NSA.
That earlier episode gave the questions surrounding the elliptic curve seeds unusual staying power. For years, researchers and privacy advocates pointed to the NSA's role in supplying the parameters as evidence that standards bodies could be influenced to embed covert access mechanisms. The absence of any public rationale for the specific seed values made the suspicion difficult to dismiss.
Claude's analysis changes the evidentiary picture by identifying a plausible, ordinary origin for the numbers. The use of «lrand48» suggests the seeds were generated with a routine tool rather than crafted through a deliberate process aimed at producing exploitable structure. The speed of the discovery, achieved in less than half an hour, also highlights how AI models are increasingly being applied to problems in mathematics and computer science that have resisted conventional analysis.
The finding is likely to be scrutinized by cryptographers, who will want to verify the derivation independently and assess whether the connection to «lrand48» holds across all the affected curves. Even a confirmed link to a simple generator would not constitute a formal security proof, and the broader question of how much trust to place in standards shaped with intelligence agency input will remain a matter of judgment.
For the institutions and developers who rely on these curves to secure digital signatures, the practical implications are limited in the short term. The standards remain in wide use, and no new vulnerability has been demonstrated. What has changed is the historical record: a long-standing suspicion now has a straightforward explanation that points away from deliberate subversion.
The episode also illustrates a shift in how cryptographic puzzles are being revisited. Problems that once required sustained human effort can now be approached with AI assistance, potentially reopening old questions and producing answers that were previously out of reach. Whether that trend strengthens confidence in existing standards or invites further scrutiny of their origins remains to be seen.
3
