An artificial intelligence model developed by Anthropic has successfully resolved one of the most enduring and controversial whodunits in modern cryptography. The AI traced a set of numbers handed over to the United States government by the National Security Agency in the late 1990s—long suspected by security experts of containing a hidden backdoor—back to an ordinary, rudimentary random-number generation program fed with computer timestamps from 1997 and 1998.
Operating inside an internal research environment, the Claude model solved in roughly half an hour a puzzle that human cryptographers, security researchers, and independent investigators had pursued for a quarter of a century. The breakthrough not only puts to rest decades of speculation regarding potential intelligence agency meddling in foundational encryption standards but also highlights the growing capability of advanced artificial intelligence models to assist in complex mathematical and historical investigations.
A Lock with an Unexplained Design
Much of the modern internet relies heavily on elliptic-curve cryptography, a sophisticated family of mathematics that underpins secure communications, digital signatures, and decentralized financial networks such as bitcoin wallets. This mathematical framework allows an individual or entity to prove they hold a secret cryptographic key without ever revealing the underlying details of the key itself.
To visualize the concept, security experts often describe the elliptic curve as a standardized lock design that anyone can inspect and implement. Each user possesses their own unique secret key, but the entire system depends on the collective trust that the underlying mathematical design is exceptionally difficult to break. However, a fundamental question has always accompanied this technology: who chose the specific mathematics that the global community would trust?
In the year 2000, the U.S. National Institute of Standards and Technology, commonly known as NIST, officially adopted a set of elliptic curves supplied by the NSA. The intelligence agency provided these specific designs to the federal standards body without fully explaining the exact origin of certain numbers used to generate them.
The aura of secrecy surrounding these choices became significantly harder for the cybersecurity community to dismiss following the historical disclosures by Edward Snowden in 2013, which linked the NSA to deliberate, multi-pronged efforts to weaken cryptographic standards and insert vulnerabilities into widespread security tools. In response to those revelations, NIST eventually removed one specific tool—the Dual EC random-number generator—from its official recommendations.
The starting numbers used to generate the curves, known to cryptographers as seeds, were public knowledge from the very beginning. Yet, the NSA consistently declined to explain how it had selected them, fueling decades of suspicion. Critics and security researchers theorized that the agency might have tested countless alternative variations until it stumbled upon a curve possessing a subtle mathematical weakness that only the NSA knew how to exploit. The prevalent theory among skeptics was that the agency had effectively baked a secret backdoor into every digital security system that adopted the standard.
Cryptographic researchers Neal Koblitz and Alfred Menezes examined the growing controversy in detail in their influential 2015 research paper, “A Riddle Wrapped in an Enigma.” Menezes, who had personally helped write the original standard years prior, recalled receiving the disputed seeds directly from NSA mathematician Jerry Solinas back in 1997.
Over the years, independent security researchers attempted to uncover the truth through various channels. Security researcher Steve Weis spent years searching for an official explanation, which included filing formal public-records requests in 2018. NIST responded by stating that it held no documentation regarding the seed origins, while the NSA largely ceased responding to inquiries. One of the more promising leads suggested that the numbers might have originated from a humorous phrase or inside joke that Solinas had since forgotten, sending investigators down lengthy paths searching through possible word combinations and linguistic patterns.
By 2023, cryptography engineer Filippo Valsorda and a collective of contributors attempted to crowdsource a solution by offering a $12,288 financial bounty. That prize pool eventually swelled to over $36,000 with the stipulation that the winnings be donated to charity. Unfortunately, Solinas had passed away earlier that year, leaving researchers to work exclusively through old recollections, historical context, and the raw numbers themselves without reaching a definitive conclusion.
Claude Goes Digging
The trajectory of the investigation shifted dramatically when Anthropic’s AI model entered the picture. Earlier this month, an employee at Anthropic tasked an internal research model with investigating the historical seeds, specifically omitting the long-standing theory that the numbers derived from a forgotten passphrase or text string.
It took the model less than half an hour to trace five distinct NIST curve seeds directly to lrand48, a standard pseudo-random number generator found natively in older C software libraries from the era. Claude successfully reproduced the mathematical sequence behind P-256—a widely deployed elliptic curve used across various secure protocols—and located an exact match for its published seed at the 1,842nd candidate generated by the software.
“This is astonishing: it solves the 25-year-old mystery,” the model noted in its recorded reasoning process after briefly pausing to evaluate whether it was operating inside a simulated test environment or dealing with historical reality. “Hmm, is it plausible in reality? As of my knowledge… the bounty was unsolved… All top bits clear—20 bits, chance 1e-6. So indeed real-world.”
Anyone with access to the historical code can now rerun the program and independently derive the exact numbers that the NSA originally supplied to NIST. According to Anthropic, this level of reproducibility renders the hypothesis of a deliberately engineered, hidden weakness far less plausible, though the research team noted that they still cannot fully explain why the program passed over certain usable candidates or how other structural parameters of the curves were selected.
That inherent predictability is precisely what allowed Claude to successfully rebuild the numbers, and experts note that such predictability would represent a catastrophic vulnerability if found inside a modern cryptocurrency wallet. A secret cryptographic key generated using a predictable legacy generator like lrand48 could theoretically be recreated by any malicious actor who managed to deduce the approximate timeframe in which it was created, granting them total access to drain the associated funds. However, because the NSA’s curve seeds were entirely public from day one, there was never a hidden secret waiting to be covertly stolen.
Where Bitcoin Fits In
Cryptocurrency networks like Bitcoin were never part of this decades-long guessing game. Bitcoin wallets operate on secp256k1, an elliptic curve built from a significantly simpler and more transparent mathematical equation.
Bitcoin’s version is defined by two small, fixed numbers that leave virtually no adjustable parameters, whereas the P-256 standard carries a 77-digit number derived from the NSA’s previously unexplained seed. With so few dials and variables available to manipulate during its creation, there has been virtually no room for anyone to have quietly embedded a structural weakness into Bitcoin’s curve, a factor that cryptographers have consistently highlighted as a major strength of the network.
Bitcoin creator Satoshi Nakamoto never publicly explained his specific rationale for choosing secp256k1, and the curve was rarely utilized in mainstream applications before Bitcoin propelled it to global prominence. Ethereum later adopted the same curve, and Bitcoin retained it even when its 2021 Taproot upgrade introduced newer styles of digital signatures.
Despite the security of its curve, the broader cryptocurrency ecosystem faces a different long-term cryptographic challenge from the horizon of quantum computing. A sufficiently powerful quantum computer would introduce an entirely different category of danger, capable of working backward from a public key visible directly on the public Bitcoin blockchain to uncover the private secret key behind it, entirely regardless of how the underlying curve was originally chosen.
Hardware advancements across the technology sector continue to narrow the gap between theoretical quantum mechanics and a practical, working attack. Google announced this week that its new Willow quantum chip had successfully reduced error rates roughly 40-fold in one of the most computationally demanding steps that such a machine would require to perform cryptanalytic tasks.
Even so, security experts emphasize that no one has yet built a quantum computer anywhere near large or reliable enough to crack a modern cryptographic key, and there remains no consensus-backed timeline for when such a device might materialize.
Meanwhile, Anthropic noted in its technical commentary that organizations and individuals currently utilizing the NIST curves do not need to alter their security practices. After a quarter-century of lingering questions, scientific suspicion, and public debate regarding the origin of the federal standards, researchers and curious developers alike can now run the software program and retrace the exact computational path that led to the published numbers.
The resolution of the mystery also brings a positive outcome for the broader digital security ecosystem. The accumulated bounty of $36,864 will be directed to the Internet Security Research Group, the non-profit organization responsible for Let’s Encrypt, which provides free digital security certificates to websites across the global web.
Comments (0)