What happened
Anthropic published findings claiming its Claude model worked out the procedure that produced the seed values for several NIST-standardized elliptic curves, according to a Bitcoinist report Sunday. The process Claude surfaced involves lrand48, a pseudo-random number generator from the standard C library that dates to early Unix, seeded with timestamps from the era. Those seeds then fed into the curve-parameter generation that NIST published in FIPS 186-2 in 2000.
The specific curves are the ones that sit on top of prime fields in the P-series, long deployed in TLS certificates, smartcards, and parts of the government crypto stack. The reconstruction does not reveal a new vulnerability. It answers a question cryptographers have asked for more than two decades: where did those arbitrary-looking seed numbers actually come from.
Why it matters
The NIST P-curves were drafted at NSA, and the agency has never published a step-by-step account of how the seeds were chosen. That silence became radioactive in 2013 after Edward Snowden's disclosures showed the NSA had engineered a backdoor into Dual_EC_DRBG, another NIST-blessed primitive. Cryptographers including Daniel J.
Bernstein and Tanja Lange have argued for years that any curve whose seed origin cannot be audited should be treated as suspect. If Anthropic's reconstruction holds up, it narrows the question. A seed produced by lrand48 and a Unix timestamp is weak by modern standards, but it is not, on its face, a trapdoor.
It is the kind of thing a 1990s engineer at a government lab would reach for because it was already on the machine. That is a very different story from a hidden parameter chosen to leak private keys.