For a quarter-century, cryptographers have eyed a set of seemingly random numbers published by the National Security Agency with deep suspicion. The constants embedded in Dual_EC_DRBG—a pseudo-random number generator the NSA pushed into federal standards in 2006—looked too arbitrary to be clean, too convenient to be coincidence. Whispers of a backdoor persisted even after Edward Snowden's leaks suggested the agency had indeed sabotaged the algorithm. This week, Anthropic's Claude chatbot did what human researchers could not: it reverse-engineered the origin of those numbers, tracing them to a mundane bug in a basic C program from the late 1990s.
The discovery is equal parts vindication and anticlimax. The constants were not the product of a sophisticated mathematical trap; they were the output of a flawed script that someone at the NSA ran once and never questioned. Claude identified the exact sequence of operations—a botched elliptic-curve point multiplication—that would produce the published values. No hidden structure, no cleverly planted weakness. Just sloppiness that happened to look sinister.
Why this matters now
Dual_EC_DRBG has been deprecated for years, removed from standards after the Snowden revelations made its use radioactive. But the episode has shaped how the security community approaches government-backed cryptography. Every new NIST standard is now scrutinized for "nothing-up-my-sleeve" numbers—constants derived from transparent, verifiable sources like the digits of π. The fact that an AI could crack a mystery that eluded expert cryptanalysts for two decades raises a different question: what else might large language models uncover in legacy code, forgotten protocols, or classified-but-leaked documents?
The timing is also pointed. Bitcoin's quantum resistance remains an open problem, and the cryptocurrency community has long debated whether early design choices—Satoshi's choice of secp256k1, the SHA-256 hash function—might harbor latent vulnerabilities. Claude's feat will not settle those arguments, but it demonstrates that AI can now interrogate cryptographic artifacts with a level of persistence and pattern-matching that human intuition struggles to match. If a chatbot can decode NSA number theory, it can certainly audit blockchain primitives.
Our take
The NSA's reputation for cryptographic sabotage is well-earned, but this case is a reminder that incompetence and malice can produce identical-looking evidence. The agency may have deliberately weakened Dual_EC_DRBG—most experts still believe it did—but the constants themselves were likely just an artifact of hurried, under-reviewed engineering. Claude's contribution is less about exonerating the NSA than about showing how AI can function as a kind of archaeological tool for digital systems, reconstructing the decisions and accidents that shaped them. That capability will matter far more in the next decade than any single solved mystery from the last one.
