Claude Solves the 25-Year-Old NSA Crypto Mystery in 30 Minutes
Anthropic used Claude to prove the NSA did not backdoor our most vital encryption standard, ending a 25-year conspiracy theory.
Research · Source: Hacker News
What happened
For 25 years, the cryptography world had a massive open question. The NSA provided the seeds for the NIST ECDSA curves in 2000 with zero explanation. People assumed the NSA hid a backdoor in the math. A popular rumor claimed NSA employee Jerry Solinas hashed a funny passphrase to make the seeds. A bounty was even placed for anyone who could find the real origin.
In October 2026, an Anthropic employee asked an internal Claude research model to investigate. In under half an hour, Claude cracked it. The NSA did not use a complex backdoor or a hashed phrase. They used a weak 1990s C library random number generator called lrand48. Claude guessed the software used an old pseudorandom number generator and wrote a scanner to test 34 combinations.
Claude noticed a pattern of zeros in the top bits of the seeds. It then used a textbook lattice problem solver with the LLL algorithm to recover the internal state. The model proved the NSA simply seeded the generator with basic Unix timestamps from May 1997 to April 1998. Claude also found that base points for some curves came from the FreeLIP long-integer arithmetic package seeded with small integers.
Key facts
- $36,864 — The bounty amount donated to the Internet Security Research Group
- 30 minutes — The time it took Claude to discover the origins of the seeds
- 1/5,810 — The probability of the top bits being zero by chance, which tipped Claude off
- 1997 to 1998 — The years of the Unix timestamps used to seed the prime curves
Why it matters
This changes how we view the security of the internet. The P-256 curve secures almost everything we do online. For decades, engineers lived with a lingering fear that the NSA had a skeleton key. Claude just proved the lock is safe. The NSA was not playing multi-dimensional chess to subvert global encryption. They were just lazy and used a basic clock to generate random numbers. We can finally trust these cryptographic standards without reservation.
The second-order effect is what this says about AI capabilities in cybersecurity and mathematics. A large language model independently hypothesized a 1990s software quirk. It wrote a custom scanner. It recognized a statistical anomaly with a one in 5,810 chance of occurring. It applied a complex basis reduction algorithm and solved a bounty that stumped human cryptographers for a quarter-century. Models are no longer just writing boilerplate code. They are executing original cryptographic forensics.
For builders
AI for legacy code forensics
Claude solved this by guessing how 1990s C libraries worked. If you maintain legacy systems, give your models the raw binaries. AI can spot patterns in archaic code that modern engineers miss.
End of the backdoor paranoia
Founders building secure messaging apps can stop worrying about the NIST curves. You do not need to roll your own crypto. The standard is safe, saving you engineering time and audit costs.
Unsolved crypto bounties remain
Claude could not solve the B-163 binary seed or the base points for P-256. Security researchers can still use AI to hunt down these remaining edge cases for prestige and profit.
My take
I build AI products every day, but this blew my mind. We spent 25 years thinking the NSA was evil, and Claude proved they were just using bad C code. If an AI can solve a decades-old cryptographic bounty in 30 minutes, human-only security research is officially dead. You either build with AI or you get left behind.
Original reporting: Hacker News. This is my rewrite and opinion.