Anthropic said on Oct. 17 that its internal model Claude reconstructed the seed generation method behind parts of the NIST standard elliptic curves in less than 30 minutes, offering a new answer to a cryptographic question that has remained unresolved for about 25 years.
The result has direct relevance for crypto infrastructure. ECDSA, or the Elliptic Curve Digital Signature Algorithm, is widely used for digital signatures and supports wallet signing and identity verification across many crypto assets. Around 2000, some parameters for NIST standard elliptic curves were provided by the U.S. National Security Agency, or NSA, but the process used to generate the seeds was never made clear, leaving room for long-running backdoor speculation.
A new explanation for how the NIST curve seeds were produced
According to Anthropic’s disclosure, Claude used reverse engineering and found that the seeds were likely generated not by a sophisticated secret method, but by combining Unix timestamps with lrand48, a pseudorandom number generator from a 1990s-era C library. That gives the industry a new explanation for the origin of parameters that had remained under question for years.
Why Bitcoin took a different path
The report ties the finding to a familiar debate in crypto. Bitcoin uses the secp256k1 curve chosen by Satoshi Nakamoto, rather than the NIST standard curves that became common elsewhere. That decision is often discussed in the context of limited transparency around standards-setting, though the report says it would be too simplistic to describe avoiding an NSA backdoor as the only reason for the choice.
This does not break ECDSA
Anthropic also draws a clear boundary around what Claude did and did not show. The model reconstructed the generation method for some standard curve parameters; it did not break ECDSA, and it did not uncover a new mathematical vulnerability in the curves involved. Reproducing the parameter-generation process may help evaluate old backdoor suspicions, but it does not amount to proof that every related implementation is free of risk.
AI’s role in cryptographic review
The test points to another use case for AI in crypto security. Beyond changing the balance in offensive and defensive research, AI can also be used to audit historical cryptographic design decisions, test security assumptions, and provide new research tools for protecting crypto infrastructure.
