An impressive breakthrough in cryptography has been achieved by an internal research model of Claude AI, developed by Anthropic. In under 30 minutes, Claude AI traced the origins of five elliptic curve seeds supplied by the NSA, resolving a question that has puzzled experts for approximately 25 years. This analysis provides clarity on how crucial public security parameters were selected without exposing secret keys or compromising encryption.
Unveiling the Source of NSA Curve Seeds
On October 16, 2026, researchers revealed that the origins of the NIST elliptic curve seeds, specifically P-192, P-224, P-256, P-384, and P-521, were rooted in lrand48, a vintage C library random-number generator initialized with Unix timestamps. This finding challenges the long-held theory that these values were hashes derived from humorous phrases.
The adoption of the Elliptic Curve Digital Signature Algorithm (ECDSA) by NIST in its FIPS 186-2 standard in 2000 was a significant milestone in digital security. While the standard detailed the seeds used by a SHA-1-based process to derive curve parameters, it lacked explanations for the origins of these seeds, leading to speculation about potential hidden weaknesses selected by the NSA. Despite the absence of any proven attacks, the origins remained a mystery, with former NSA mathematician Jerry Solinas previously hinting at possible hash inputs like humorous phrases.
Claude’s Methodical Approach to Seed Analysis
Moving away from the humorous phrase hypothesis, Claude AI focused on generators prevalent in 1990s software. It crafted a scanner that explored 34 combinations of generators and output slices to identify patterns within the published seeds. A notable clue was found in the binary-curve seeds, where their 32-bit words consistently displayed a zero in the highest bit, aligning with lrand48’s output.
For the prime-curve seeds, which only retained eight bits from each generator output, Claude AI employed the Lenstra–Lenstra–Lovász lattice reduction algorithm (LLL) to deduce internal states from the partial values. This process confirmed that all five NIST seeds and five ANSI X9.62 example seeds were constructed similarly. By reversing the generator, initialization values corresponding to dates between May 1997 and April 1998 were uncovered, with the P-256 seed being identified as the 1,842nd candidate from August 16, 1997.
Implications and Future Perspectives
Anthropic emphasizes that this discovery does not constitute a cryptographic break or indicate any new vulnerabilities in ECDSA. The reconstructed seeds are public curve-generation data and do not allow attackers to forge signatures. The research underscores that the simplicity and repeatability of the process make deliberate manipulation unlikely.
This achievement exemplifies AI’s role in advancing cryptographic research, offering historical insights rather than new attacks. Claude AI also traced the base-point generation for P-192 and several ANSI example curves to FreeLIP, a long-integer arithmetic package. However, not all mysteries have been solved, as the seed for B-163 and several base points remain unexplained.
Anthropic has shared reproduction code for verification purposes, and the Internet Security Research Group will receive a $36,864 bounty. Although the findings do not necessitate changes to existing systems, they enhance trust in the selection of these cryptographic parameters.
