For 25 years, cryptographers wondered where the NSA got the "seed" numbers behind NIST's elliptic curves, including P-256, which secures much of the internet. In October 2026, Anthropic reported that an internal Claude research model traced them, in under 30 minutes, to an old random number generator called lrand48 fed with timestamps from 1997 and 1998. Anthropic says this makes a hidden backdoor much less plausible and users need do nothing.
Key facts: Announced: October 16, 2026, by Anthropic's Frontier Red Team. Mystery: the unexplained seeds behind NIST curves P-192, P-224, P-256, P-384 and P-521. Answer: lrand48 output seeded with Unix timestamps from May 1997 to April 1998. P-256 match: candidate 1,842. Time: under half an hour. Bounty: $36,864, donated to the Internet Security Research Group. Action needed: none.
What was the 25-year NSA number mystery?
The NIST elliptic curves, standardised around 2000, are used in TLS, digital signatures and many other security systems. Each curve was generated from a "seed" value. The NSA supplied those seeds without explaining where they came from.
That gap fed suspicion for decades. If someone chose the seeds on purpose, they might have picked curves with a hidden weakness. Cryptographers Neal Koblitz and Alfred Menezes explored the puzzle in their paper A Riddle Wrapped in an Enigma. In 2023, cryptographer Filippo Valsorda launched a public bounty for whoever could explain the seeds.
What did Claude discover about the NIST curve seeds?
According to Anthropic's research post, the seeds were not random hash outputs at all. They came from lrand48, a basic random number generator in older C libraries, seeded with Unix timestamps from May 1997 to April 1998.
In simple terms, someone at the time likely used the clock as a starting point for a standard random function. Anthropic's own summary: "It's just timestamps."
How did Claude find the answer?
An Anthropic researcher asked an internal research model to investigate, without giving it earlier theories, according to CoinDesk. The model:
- Tested hypotheses about how 1990s software might have produced the seeds.
- Tried lrand48 with timestamps from the period.
- Reproduced the sequence behind P-256 and found an exact match at the 1,842nd candidate.
- Did it all in under 30 minutes.
CoinDesk reports that the model's own recorded reasoning called the result "astonishing." The work was credited to ten Anthropic researchers, including Nicholas Carlini and Steve Weis. Weis had studied the seed question before, as described on his blog.
Does this mean the NSA did not backdoor the curves?
It makes a backdoor much less likely, but it does not prove a negative. Anthropic says the finding makes deliberate manipulation of the seeds much less plausible, and that it is not a cryptographic break or weakness.
Some questions remain open, according to Anthropic:
- Skipped candidates: candidate 876 in the same run also produces a valid curve, so it is unclear why some usable candidates were passed over.
- B-163: one binary-curve seed is still unexplained.
- Base points: how the base points for several curves were chosen is still unknown.
Do I need to do anything if I use P-256?
No. Anthropic says there is no action to take. P-256 and the other NIST curves are as secure as they were before. The discovery explains their history; it does not change their maths.
One lesson for developers: lrand48 is not a safe source of randomness for keys. CoinDesk notes that keys generated with lrand48 would be predictable, a serious risk in a crypto wallet.
What does this mean for Bitcoin?
Nothing changes for Bitcoin. It uses a different curve, secp256k1, which has fewer adjustable parameters and was never part of this mystery. CoinDesk notes the bigger long-term worry for Bitcoin is a future quantum computer powerful enough to break today's keys, which does not exist yet.
Who gets the $36,864 bounty?
The bounty Valsorda started in 2023 totals $36,864. Anthropic says it will be donated to the Internet Security Research Group, the nonprofit behind Let's Encrypt.
Why is this a big deal for AI?
It is one of the clearest public examples of AI doing real research. The model did not just summarise existing work. It formed a hypothesis, tested it and found an answer that experts had sought for years.
That fits a wider trend this month. AI tools are reverse engineering software, as with the viral REA tool, and solving long-standing technical puzzles. Expect more "AI solved it" headlines, and more scrutiny of whether each claim holds up.
What can marketers learn from this story?
The story spread because it had everything: a 25-year mystery, the NSA, a precise answer and a quotable moment. It also came with clear evidence and a clear "you don't need to do anything" message, which kept it from turning into panic.
For brands, that is the model for technical communication: lead with the answer, show the proof, and tell people what it means for them. AI assistants increasingly favour content written that way. Our GEO guide explains how to structure content so AI cites it, and MarkGrid's Model Share shows how often ChatGPT, Gemini, Perplexity, Claude and Copilot recommend your brand.
FAQ
What did Claude discover about the NSA curves?
An Anthropic research model found that the seeds of the NIST elliptic curves came from the lrand48 random number generator, seeded with timestamps from 1997 and 1998.
Does this mean NIST P-256 is broken?
No. Anthropic says this is not a cryptographic weakness and users do not need to take any action.
How long did it take Claude to solve the mystery?
Under half an hour, according to Anthropic.
Does this affect Bitcoin?
No. Bitcoin uses the secp256k1 curve, which was not part of the NIST seed mystery.
Who started the bounty for the NIST curve seeds?
Cryptographer Filippo Valsorda started the bounty in 2023. The $36,864 prize is going to the Internet Security Research Group.
Is the backdoor theory now disproven?
Not fully. Anthropic says a backdoor is now much less plausible, but some details, like skipped candidates and base points, are still unexplained.
Last updated October 21, 2026.
How often MarkGrid is named when AI models discuss this topic. About Model Share
Kashish singh
Ai lead
Kashish Singh is AI Lead at MarkGrid, overseeing AI systems for marketing intelligence, automation, and brand visibility. Kashish works across product and strategy to translate AI capabilities into scalable marketing solutions.
See it on your brand
Thirty minutes on where marketing is leaking for you.

