Anthropic Says Claude Solved NIST Curve Mystery, Finds No Backdoor

The Puzzle

For over two decades, a set of numbers has bothered cryptographers. These numbers are the seeds for the elliptic curves that secure much of today’s digital communications. The US National Institute of Standards and Technology (NIST) standardized these curves in 2000. The NSA supplied the seed numbers without explaining how they were generated. That secrecy fed a long-running theory: the NSA deliberately weakened the encryption.

Anthropic says its AI model, Claude, has cracked the mystery. According to the company, Claude traced the key numbers to a simple random-number generator seeded with timestamps. The finding eases concerns that the NSA built a backdoor into these standards. Anthropic reports no security vulnerability was found. Users do not need to take any action.

The Backstory

The NIST curves are everywhere. They are in TLS certificates, SSH keys, and many other protocols. The seeds are the starting points for the math that defines the curves. Without knowing how the seeds were picked, you cannot fully audit the curves for hidden weaknesses. The NSA’s refusal to explain the process looked suspicious.

The suspicion was not baseless. The NSA has a history of weakening crypto standards. The Dual_EC_DRBG random number generator, also standardized by NIST, was later shown to have a potential backdoor. So the curve seeds drew intense scrutiny.

The common theory was that the seeds came from a secret phrase or a hashed message. People searched for that phrase for years. Some thought it was a quote from a movie or a line from a book. None of that panned out.

How Claude Solved It

Anthropic says Claude did not search for a secret phrase. Instead, it analyzed patterns in the numbers themselves. The model noticed that the seeds looked like outputs from a specific type of random-number generator. That generator is called lrand48, a common function in old C libraries.

lrand48 is not a cryptographically secure generator. It is a simple linear congruential generator. Its output depends on a 48-bit internal state. If you know the state, you can predict the next number.

Claude reconstructed the internal state from the published seeds. The recovered starting points correspond to Unix timestamps from 1997 and 1998. Reproducing the original process generated seed values that match the published ones exactly. The whole thing took under 30 minutes of computation.

What This Means

If the seeds came from a simple random-number generator with known timestamps, that is not a sign of a backdoor. It is a sign of a shortcut. The NSA likely used whatever tool was handy in the late 1990s to generate the numbers. They did not bother to document it.

Anthropic says this weakens the deliberate backdoor theory. A deliberate backdoor would require careful construction of the seeds to create a hidden weakness. Using lrand48 with timestamps does not fit that pattern. It looks more like an engineering decision, not a malicious one.

But the company adds a caveat. The assessment is based on Anthropic’s account. There is no reported independent verification. The findings are Anthropic’s own claims. No third-party review is mentioned in the reporting.

Not Everything Is Solved

Anthropic did not explain every number. Several curve parameters remain unresolved. One seed is still unexplained. The base points of the four main NIST curves are also not fully traced. The original process skipped some seemingly usable candidate curves. That means the full history is not yet known.

Anthropic says it found no evidence of deliberate weakening. But the lack of a full explanation means some questions remain. The skipped curves are curious. Why would the NSA pass over valid candidates? The company does not offer a definitive answer.

The Lenstra Connection

The model also traced the starting points of six older curves. Those curves came from a software package by Arjen Lenstra, a well-known math software developer. Lenstra’s package used a random-number generator to produce those seeds. That is a separate finding from the NIST curve work.

This adds context. It shows that using simple generators for curve seeds was not unusual. The practice was common in the field. It makes the NIST seeds look less like an anomaly.

No Action Needed

Anthropic is clear: there is no vulnerability. The curves are safe to use. The mystery was about the origin of the numbers, not about a flaw in the math. Users do not need to change anything.

The finding is more about history than about security. It closes a chapter of speculation. The backdoor theory was a persistent shadow over these standards. If Anthropic is correct, that shadow is gone.

The Caveat

The reporter’s note is important. The assessment that this weakens the backdoor theory is based on Anthropic’s account. There is no independent verification. The reporting mentions no third-party review.

This is not unusual for AI company announcements. The model’s reasoning is complex. Verifying it requires deep expertise in both cryptography and the history of the standards. That takes time.

For now, the crypto community has a new explanation to consider. It is plausible. It fits the data. But it is not yet confirmed by outside experts.

The Bigger Picture

This is a case of an AI model doing what it does best: finding patterns in large amounts of data. The seeds were a small set of numbers, but the pattern was hidden. Claude found it by looking at the structure, not by guessing.

The result is a practical outcome. It resolves a long-standing question. It also shows a use case for AI in security research. Not as a tool for breaking encryption, but as a tool for understanding how encryption was built.

Anthropic’s claim is significant. If it holds up, it changes the narrative around NIST curves. It turns a story of suspicion into a story of mundane engineering. That is a less exciting story, but a more likely one.

The unresolved parameters remain. The skipped curves remain unexplained. But the core mystery — where did the seeds come from — may now have an answer.