Claude Solves a 25-Year-Old Cryptography Puzzle in About 30 Minutes
Anthropic says Claude traced the mysterious seed numbers behind NIST's standard elliptic curves to a simple 1990s random number generator. It is a neat result, but not a security break.
Anthropic reports that Claude has explained where the “seed” numbers behind several widely used encryption curves came from, a question that has nagged cryptographers for about 25 years. According to the authors, it took Claude roughly 30 minutes. The answer is almost comically ordinary: an old random number generator from the C programming language, started with a clock reading.
Some background. In 2000, the US standards body NIST published a set of elliptic curves, which are the mathematical shapes behind everyday encryption such as the padlock in your browser. Each curve was built from a published seed, a starting number, supplied by the NSA without explanation. Without a reason, people wondered whether the seeds hid a deliberate weakness, a so-called backdoor. A popular story said the seeds were a hash (a one-way scramble) of a joking passphrase. Anthropic’s team says that story is wrong.
Instead, the seeds for the curves P-192, P-224, P-256, P-384 and P-521 came from a generator called lrand48, started with a Unix timestamp (the computer’s clock, counted in seconds). For P-256 the program simply tried candidates until the 1,842nd one passed the checks. The timestamps fall between May 1997 and April 1998. Claude also explained where the base points for six other curves came from: a random number generator in an old maths library called FreeLIP.
How did it get there? Claude guessed the seeds came from a weak 1990s generator, then scanned for tell-tale patterns. Four seeds had a zero in the top bit of all 20 words, which would happen by chance about once in 5,810 times. That matched how lrand48 behaves. Recovering the hidden internal state took a mix of brute force and a lattice method called LLL. The team published code anyone can run to reproduce the seeds.
What’s behind it: a backdoor needs the creator to choose the numbers cleverly. If the seeds come from a clock and a boring generator, the backdoor theory gets much harder to defend. It is also a good example of AI as a patient detective: it tests many hypotheses quickly and a human checks the evidence.
Worth keeping expectations grounded. The authors say this is not a break of ECDSA or these curves. Some questions remain open, including other base points and one binary-curve seed. They also admit that someone could, in theory, have tried many timestamps and kept a favourable result, and nothing in the data rules that out. Anthropic is donating the $36,864 bounty to the Internet Security Research Group.
What this means for you: nothing changes in how you browse or bank. Your encryption is no less safe today than yesterday. The takeaway is that AI is getting good at old, stubborn puzzles where the answer can be checked, which is a promising sign for security research.
Sources
Source: https://www.anthropic.com/research/ecdsa-curve-origins
DeepSeek V4.1 Flash 1017: Cheaper Agent Work and a Search API You Can Rent
DeepSeek released an open-weights update aimed at agent and coding tasks, claims about $0.20 per task, and opened its search engine to outside developers at $0.75 per 1,000 calls.
