Claude Solves 25-Year-Old NSA Cryptography Mystery in 30 Minutes

Anthropic's AI model solved a 25-year-old NSA encryption puzzle, proving global security standards rely on a basic 1990s computer clock.

Models · Source: Escudo Digital

What happened

For 25 years, cryptographers suspected the NSA hid a backdoor in global encryption standards. The NIST FIPS 186-2 standard used unexplained numerical seeds to define security elliptic curves. These curves form the technical foundation for much of the digital encryption used worldwide. Conspiracy theories claimed these numbers were secret phrases or malicious backdoors designed to spy on protected communications. Anthropic asked an internal Claude research model to investigate the origins of these seeds without steering it toward the old theories.

Claude solved the puzzle in under 30 minutes. The AI ignored the secret phrase theories and looked for patterns matching 1990s software. It noticed that certain high bits in the seed values were consistently zero. This signature pointed directly to lrand48, a standard linear congruential pseudorandom-number generator from old C programming libraries. The NSA merely activated this generator using basic Unix timestamps recorded between May 1997 and April 1998.

Claude used the Lenstra-Lenstra-Lovász lattice reduction algorithm to run the generator backward and recover the internal states. The AI wrote a custom scanner to test combinations of older generators and output slices. The seed for the P-256 standard perfectly matched the 1,842nd candidate generated from a run started on August 16, 1997. The discovery claims a $36,864 bounty posted by cryptographers in 2023. The reward will be donated to the nonprofit Internet Security Research Group.

Key facts

Why it matters

This proves the foundational mathematics of our digital security are safe from intentional NSA sabotage. By demonstrating that the figures originated from a mundane computer clock routine, the probability of a hidden espionage mechanism is drastically reduced. You do not need to rotate keys, replace certificates, or update systems. The encryption itself remains unbroken, and private keys were never exposed.

This marks a massive shift in how we use AI for historical and mathematical forensics. Claude did not just guess passwords or read documentation. It analyzed mathematical structures, wrote custom code, and applied complex lattice reduction algorithms to reverse-engineer decades-old software states. AI is now capable of solving complex cryptographic provenance issues that human experts debated endlessly. Builders can now trust AI to audit the fundamental building blocks of legacy infrastructure.

For builders

No need to rotate your encryption keys

This discovery is historical reconstruction, not a vulnerability. Your ECDSA implementations remain entirely secure. No one pays a penalty for this finding, and you do not need to migrate your cryptographic systems or revoke active certificates.

AI excels at legacy code forensics

Claude solved this by testing 34 combinations of older generators and output slices. Builders can deploy large language models to reverse-engineer undocumented legacy systems or trace the provenance of obscure codebases. Companies paying expensive consultants for system archaeology stand to save massive amounts of time and capital.

Unexplained cryptographic mysteries remain open

Claude did not solve every piece of the puzzle. The B-163 seed and the base points for P-224, P-256, P-384, and P-521 are still unexplained. Security researchers and AI builders have a clear opportunity to build specialized models to crack these remaining historical anomalies.

My take

Human experts spent a quarter of a century chasing conspiracy theories about NSA backdoors. Claude looked at the math, ignored the drama, and found a lazy developer using a basic C library clock in under 30 minutes. This is exactly why we need AI to audit legacy systems. It does not care about human paranoia. It just runs the numbers and finds the truth.

Original reporting: Escudo Digital. This is my rewrite and opinion.

More AI news for builders