The Frontier Red Team at Anthropic announced on October 16, 2026, that its Claude model uncovered how the seed parameters behind the U.S. National Institute of Standards and Technology (NIST) ECDSA elliptic curve cryptography standard were generated. This finding closes a debate spanning more than a quarter of a century within the cryptography community regarding whether the parameters provided by the U.S. National Security Agency (NSA) in 2000 contained intentional vulnerabilities.
According to the research paper, it took Claude less than half an hour of independent analysis to determine that the seed values for prominent curves such as NIST P-192, P-224, P-256, P-384, and P-521 did not originate from complex hashing algorithms or secret passphrases. Instead, they were simply the outputs of the classic C standard library pseudorandom number generator lrand48, seeded (srand48) using Unix timestamps ranging from May 1997 to April 1998.
Methodology and Empirical Evidence
This mystery had endured across multiple generations of experts because previous hypotheses focused on searching for whether the NSA had hashed a humorous passphrase. Claude approached the problem from a different angle by reconstructing the programming practices of engineers in the 1990s, conjecturing that parameter-generation software from that era typically relied on built-in system random number generators.
By writing a bit-level scanner to analyze the standard seeds, the model identified an unmistakable fingerprint: the most significant bits in the 32-bit words of the sequences were all zeroes, a hallmark characteristic of the lrand48 linear congruential generator algorithm. Subsequently, applying the Lenstra–Lenstra–Lovász (LLL) lattice basis reduction algorithm, the model reversed the generator’s state back to its starting point and precisely reproduced all 160 bits of the published parameters, with an odds of random coincidence being practically zero.
Strengthening Confidence in Cryptographic Standard Security
Anthropic emphasized that this finding is by no means a cryptographic break or the discovery of a new security vulnerability. On the contrary, proving that the parameters were selected pseudorandomly based on a computer clock relieves the cybersecurity community of concerns regarding subtly planted backdoors. Researchers can have greater confidence that the curves securing much of today’s internet traffic and digital assets were constructed through a straightforward random process, free from covert manipulation.
Nevertheless, open questions remain. The seed parameter for the binary curve B-163 still does not fit this pattern, and the base point selection method for larger curves like P-256 or P-384 has not been fully illuminated. Still, an AI system autonomously piecing together historical clues and classical algorithms to resolve a decades-old problem demonstrates that the role of large language models extends far beyond routine code generation, opening new avenues for auditing critical legacy technologies.