Claude Traces NIST Curve Seeds to a Random Number Generator

Anthropic says its Claude research model has resolved a decades-old question about the origins of cryptographic parameters used in widely deployed digital signature standards. In October 2026, the model traced the seeds for five NIST elliptic curves to a conventional random number generator initialized with a Unix timestamp, offering new evidence against concerns that the parameters were deliberately designed with hidden weaknesses.

According to Anthropic, the discovery covers the NIST P-192, P-224, P-256, P-384 and P-521 curves. The investigation also identified the same generation pattern in five additional seeds from the ANSI X9.62 standard. The company emphasized that the findings do not demonstrate a cryptographic break, a new attack or a vulnerability in the curves.

The Elliptic Curve Digital Signature Algorithm (ECDSA) is widely used for digital signatures. NIST adopted the relevant elliptic curve specifications in 2000 as part of the FIPS 186-2 Digital Signature Standard, which included curves with different sizes and security levels.

The specifications included seed values supplied by the National Security Agency (NSA), but did not explain how those values had been generated. That omission fueled years of speculation about whether the parameters could have been manipulated to introduce hidden weaknesses.

Anthropic said its findings instead point to lrand48, a pseudorandom number generator included in C libraries, initialized using a Unix timestamp. The method repeatedly generated candidate byte sequences until it produced the required standardized seed.

For the P-256 curve, the process required 1,842 iterations to find the published seed, according to Anthropic.

Anthropic said the results increased its confidence that the curves were not deliberately weakened during their creation. The company also said the findings do not require users to change their existing practices.

The investigation began after previous attempts to identify the seeds' origins failed. One longstanding theory suggested that Jerry Solinas, a longtime NSA employee, had derived them by hashing a humorous phrase. That explanation remained unverified, and Solinas died in 2023.

Anthropic said an employee revisited the problem in October 2026 by asking Claude to investigate the seed-generation process without directing it toward the earlier theory.

Rather than searching for a phrase, Claude examined how software from the period might have generated the values. It developed a scanner to compare portions of the published seeds against consecutive outputs from pseudorandom number generators associated with 1990s-era C libraries.

The analysis examined 34 combinations of generators and output slices. Claude then identified a bit pattern consistent with lrand48, which returns only the upper 31 bits of its 48-bit internal state.

Anthropic reported that four seeds displayed a distinctive pattern in their highest bits. Claude recovered the generator states by testing possible hidden bits and checking whether subsequent outputs matched the published values.

For the remaining prime-curve seeds, the model used a mathematical technique involving the Lenstra–Lenstra–Lovász (LLL) lattice basis reduction algorithm to recover the internal state from truncated outputs.

According to Anthropic, the recovered states reproduced all 160 bits of the relevant seeds. The same output pattern appeared across the five NIST prime-curve seeds and five additional ANSI X9.62 seeds, supporting the identification of the generator.

The company said the discovery provides a reproducible explanation for how the seed values were produced, addressing a longstanding uncertainty surrounding the origins of these cryptographic parameters.

Follow suaragarut.id Add to preferred sources on Google

• Artificial Intelligence
• Cybersecurity
• Cryptography