Anthropic Says Claude Cracked a 25-Year NSA Cryptography Seed Mystery in 30 Minutes
An artificial intelligence model built by Anthropic has apparently untangled a cryptographic riddle that has puzzled security researchers for a quarter of a century. The company said its Claude system traced the origin of unexplained numbers embedded in a widely used federal encryption standard back to a mundane software function seeded with Unix timestamps from the late 1990s.
The finding, which Anthropic said took Claude less than 30 minutes to produce, addresses long-standing suspicion that the National Security Agency may have deliberately weakened the elliptic curve parameters it supplied when the National Institute of Standards and Technology standardized them in 2000. Anthropic said the results do not expose a security flaw and require no action from users.
The mystery revolves around the Elliptic Curve Digital Signature Algorithm, or ECDSA, which verifies the authenticity of digital messages, software and transactions. For roughly 25 years, cryptographers have questioned how the NSA generated specific constants used in the curves. Because the agency never disclosed its derivation method, some researchers speculated the numbers could have been chosen to create a backdoor.
Those concerns were not purely theoretical. The broader cryptography community had previously identified a potential backdoor in Dual_EC_DRBG, another government-linked algorithm. That episode deepened skepticism about unexplained parameters in NSA-influenced standards, even in the absence of proof of intentional sabotage.
Claude approached the problem differently from earlier attempts. Rather than hunting for a rumored secret phrase, the model analyzed the numbers for statistical patterns. It identified lrand48, a random-number generator commonly found in older C programming libraries, and reconstructed its internal state. The recovered starting points corresponded to Unix timestamps from 1997 and 1998.
By reproducing the original process, Claude generated the published seed values exactly. Anthropic said that level of precision offers strong evidence of how the numbers were created. The model also linked the starting points of six older curves to a random-number generator in a mathematical software package developed by Arjen Lenstra.
Not every question has been answered. The researchers could not explain the origins of several other curve parameters, including one seed and the base points of four major NIST curves. They also found that the original process skipped some candidates that appeared capable of producing valid curves, though they found no evidence the selections were deliberately weakened.
Anthropic said the findings make the theory of an intentional backdoor less plausible. The company framed the work as an example of AI systems being applied to long-standing problems in mathematics and security research, where pattern recognition at scale can uncover relationships that manual analysis might miss.
The development arrives amid broader debate over the trustworthiness of foundational cryptographic infrastructure. Blockchain networks, custody systems and encrypted communications all depend on primitives whose integrity underpins user confidence. Renewed scrutiny of legacy standards tends to attract attention from security researchers monitoring those assumptions, even when the standards are not directly tied to digital assets.
The reported breakthrough also highlights a separate but related conversation about how AI compute and traditional cryptographic mining allocate their costs.
Vitalik Buterin, co-founder of Ethereum, published a comparison on October 17 that distinguished the economics of Bitcoin mining from AI inference. He described Bitcoin mining costs as split roughly evenly between computing capital expenditure and electricity, while AI inference is overwhelmingly dominated by hardware investment.
The mining table Buterin shared placed electricity at 45.9% of the estimated cost per Bitcoin. Direct power to computers accounted for $40,600 per BTC, or 43.3%, with other electricity expenses adding $2,400, or 2.6%. Computing capital expenditure contributed approximately $37,600 per BTC, or about 40%, covering hardware, cooling, power equipment and financing. Labor, rent and other expenses made up the remainder of the roughly $93,700 per BTC total.
The AI table used a different unit, dollars per GPU-year, and assigned approximately 76% of total costs to computing capital expenditure. That category came to about $19,900 per GPU-year. Electricity supplied directly to computers was just $1,050 per GPU-year, or 4%, with other electricity adding $210, or 0.8%. Labor, rent and other operating expenses filled out the estimated $26,260 per GPU-year total.
The distinction matters because the two industries are frequently lumped together in debates over energy consumption and environmental impact. Buterin's breakdown suggests that comparison oversimplifies how each sector spends money and responds to cost pressures. A spike in electricity prices hits Bitcoin miners harder, since power is their dominant recurring expense. AI operators face greater exposure to the price and availability of advanced chips, because capital investment in hardware represents the larger share of their costs.
That dynamic is already visible in how data center operators structure their deals. TeraWulf, for instance, separated AI hosting from processor ownership in its 20-year lease with Anthropic at the Justified Data campus in Kentucky. Under that arrangement, customers remain responsible for servers, processors and technology upgrades, while TeraWulf supplies power and physical infrastructure. The facility is expected to support roughly 401 megawatts of critical computing capacity, with initial capacity anticipated in the second half of 2027.
The AI-assisted cryptography discovery does not carry a direct market event, since it concerns historical government standards rather than blockchain protocols or digital assets. Still, it may renew broader interest in how trustworthy foundational cryptography is, a topic that resonates with crypto markets given their reliance on secure algorithms. The use of AI to analyze such problems may also draw attention to AI-assisted security auditing as an emerging tool within the broader cybersecurity and digital asset industries.
Further technical detail and independent verification will likely shape how the finding is ultimately interpreted. The available reporting did not indicate whether the NSA has commented on the claim.