Why this matters: an AI model just explained the unexplained numbers behind widely used digital signatures — and the plain answer makes a hidden backdoor far less likely.
October 18, 2026 — Anthropic says an internal Claude research model has traced the mysterious seeds behind five NIST elliptic curves to an ordinary 1990s random-number generator, solving in under 30 minutes a question open for about 25 years.
Image: A glowing elliptic curve over a vintage computer clock and streams of numbers, symbolising the solved seed mystery. Credit: AI-generated illustration / Red Eye Eagle News, original illustration.
Editor's Highlight: The seeds were not a hidden phrase. According to Anthropic, they were output from lrand48, a simple generator in older C libraries, started from a computer clock in 1997 and 1998 — a process anyone can now rerun and check.
Table of Contents
- Key Facts
- A Mystery That Lasted 25 Years
- How Claude Found the Fingerprint
- What the Answer Does and Does Not Prove
- Why It Matters Beyond One Set of Curves
- Bottom Line
- FAQ
- Sources
Key Facts
| Detail | Information |
|---|---|
| Who reported it | Anthropic, in a research report published October 16, 2026 |
| What was solved | Origin of seeds for NIST curves P-192, P-224, P-256, P-384 and P-521 |
| How long it took Claude | Under 30 minutes, according to Anthropic |
| Generator identified | lrand48, from older C libraries, seeded with a Unix timestamp |
| Key match | P-256 seed was candidate 1,842 from a run starting August 16, 1997, at 01:52:34 UTC |
| Security effect | No break, no new weakness, and no action required, Anthropic says |
| Bounty | $36,864, to go to the Internet Security Research Group |
A Mystery That Lasted 25 Years
Much of the internet relies on elliptic-curve cryptography, which proves a party holds a secret key without revealing it.
These curves are used with ECDSA, the Elliptic Curve Digital Signature Algorithm, which verifies who signed data and whether it changed.
In 2000, the National Institute of Standards and Technology, or NIST, adopted the curves in its FIPS 186-2 standard. The designs had been supplied by the National Security Agency, or NSA.
The standard published seeds — starting numbers used to derive curve parameters — but never explained where they came from.
Newsroom Focus: A public seed is not a private key. Rebuilding it cannot forge a signature, but it does show whether the starting point can be trusted.
That gap fuelled suspicion that the NSA had searched for a curve with a weakness only it understood.
Former NSA mathematician Jerry Solinas had recalled a humorous phrase, along the lines of "Give Bob and Jerry a raise." That lead sent earlier hunters searching word combinations, with no result.
In 2023, engineer Filippo Valsorda and others offered a $12,288 bounty, rising above $36,000 if donated to charity. Solinas died that year.
How Claude Found the Fingerprint
According to Anthropic, an employee asked an internal model to investigate, without steering it toward the forgotten-phrase theory.
Claude tried a different idea: late-1990s software probably used a simple C-library generator, and simple generators leave fingerprints.
It built a scanner covering 34 combinations of generators and output slices, Cybersecurity News reports. The clue was in the bits: some seeds showed a repeated pattern in which the highest bit of each 32-bit word was zero.
That points to lrand48, which returns the top 31 bits of a 48-bit internal state. A clear top bit is exactly the kind of trace it can leave.
The prime-curve seeds were harder, because they kept only eight bits from each output. For those, Claude used the LLL lattice reduction algorithm to recover the internal state from partial values.
Running the generator backward revealed the starting values. Each corresponded to a Unix timestamp between May 1997 and April 1998.
Analysis Note: The strongest part of this claim is that it is reproducible. Anthropic published code so others can rerun the generator from those timestamps and compare the result with the published seeds.
| Date | Event |
|---|---|
| May 1997 – April 1998 | Timestamp window for the recovered prime-curve seed runs |
| August 16, 1997 | P-256 run starts at 01:52:34 UTC; published seed is candidate 1,842 |
| 2000 | NIST adopts the curves in FIPS 186-2 |
| 2023 | Valsorda and contributors offer the seed bounty |
| October 16, 2026 | Anthropic publishes the Claude finding |
What the Answer Does and Does Not Prove
Anthropic stresses a clear limit. This is not a cryptographic break, not a new attack, and not a newly discovered weakness in the ECDSA curves.
The seeds were public from the start and are not secret signing keys.
A simple, repeatable process leaves much less room for a hidden search for a weakened curve.
Ripple head of engineering J. Ayo Akinyele called the result remarkable on X, according to U.Today, saying it offers reproducible evidence and greater confidence about the standards' origins.
The account is still incomplete. The seed for B-163 remains unexplained, as do the base points for several curves, and researchers found skipped candidates, including P-256 candidate 876, without knowing why they were passed over.
Practical Takeaway: If you use systems built on these curves, the sources agree on one point — there is nothing to rotate, patch or change because of this discovery alone.
Why It Matters Beyond One Set of Curves
This is also a story about useful AI research. Claude did not break a modern cipher. It searched a historical puzzle in a new direction and produced a checkable answer.
For everyday readers, the lesson is simple: shared standards deserve a public paper trail. When it goes missing, suspicion can last for decades.
Bottom Line
Anthropic says Claude found that five mystery NIST seeds came from lrand48 seeded by a 1997–98 clock. It does not break ECDSA, requires no action, and makes a deliberate backdoor far less plausible. Watch for independent reruns.
FAQ
What exactly did Claude discover?
Claude found the seeds for P-192, P-224, P-256, P-384 and P-521 were lrand48 output started with 1997–98 timestamps, not hashes of a forgotten phrase, Anthropic says.
Does this mean ECDSA was broken or backdoored?
No. Anthropic says this is not a break or new weakness. The simple, reproducible process makes a deliberate backdoor much less plausible.
Do I need to change anything I use?
No action is required because of this discovery alone, according to Anthropic. The seeds are public generation values, not your private keys.
Why did people suspect the NSA for so long?
The NSA supplied the curves without explaining the seeds, and later disclosures deepened distrust. With no documentation, the question stayed open.
What happened to the bounty?
The $36,864 bounty will go to the Internet Security Research Group, the nonprofit behind Let's Encrypt, according to Anthropic's account as reported by CoinDesk and Cybersecurity News.
Sources
- CoinDesk — Anthropic's Claude cracks a 25-year NSA curve mystery, October 17, 2026: https://www.coindesk.com/tech/2026/10/17/wow-it-s-a-real-discovery-claude-cracks-a-25-year-nsa-number-mystery
- Cybersecurity News — Claude AI solves 25-year NSA cryptography mystery in just 30 minutes, October 17, 2026: https://cybersecuritynews.com/claude-ai-cryptography-mystery/
- U.Today — Ripple's head of engineering reacts to AI solving decades-old cryptography mystery, October 18, 2026: https://u.today/ripples-head-of-engineering-reacts-to-ai-solving-decades-old-cryptography-mystery
- Image — AI-generated illustration / Red Eye Eagle News, original illustration
Join the conversation
Comments
Post a Comment