Photo: Merlin Lightpainting / Pexels
Two researchers cracked increasingly large semiprimes in 16 days using AI tools, raising quiet but serious questions about encryption's long-term durability.
Cryptographers have been chipping away at RSA factoring records for decades, but the pace just got uncomfortably fast. On September 19, 2026, researcher Stephen A. Weis factored RSA-896, a 270-decimal-digit semiprime, with help from Anthropic’s Claude AI model. The achievement moved the public factoring record from 862 bits to 896 bits in 16 days.
To put that timeline in perspective: the previous record had been set on September 3, 2026, by Eric Lu of Cognition, who used the company’s Devin AI agents to factor the 862-bit RSA-260 number. Two researchers, two different AI systems, two records in just over two weeks.
What actually happened here #
Both breakthroughs relied on the general number field sieve, or GNFS, the classical algorithm that has dominated large-number factoring since the early 1990s. No quantum computing was involved. The AI assistance accelerated and optimized the classical approach rather than replacing it with something exotic.
Lu’s RSA-260 factoring consumed roughly 4,900 GPU-days at an estimated compute cost of $400,000. Weis’s RSA-896 effort has not had its compute cost published, but the 16-day gap between the two records suggests the AI-assisted optimization is meaningfully compressing what used to be multi-year projects.
The RSA Factoring Challenge, which originally sponsored these efforts, ran from 1991 to 2007. The $75,000 prize that once accompanied RSA-896 is long gone. Weis and Lu were chasing records, not prize money.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Why the RSA challenge numbers matter #
RSA encryption works because multiplying two large prime numbers together is easy, but reversing the process, finding those original primes from the product, is extraordinarily hard. The security of RSA-2048, the standard most widely deployed today, depends on that difficulty remaining computationally intractable.
RSA-896 is not RSA-2048. The gap between 896 bits and 2048 bits is not linear. Each additional bit roughly doubles the difficulty, so 2048-bit numbers remain orders of magnitude harder to factor than anything broken so far.
The previous major milestone, RSA-250 at 829 bits, was factored in 2020 using purely classical methods. Going from 829 bits in 2020 to 896 bits in 2026 is a meaningful jump, and the acceleration introduced by AI tooling suggests the curve may steepen further.
What this means for cryptographic infrastructure #
For organizations running systems that depend on RSA encryption below 2048 bits, these records are a concrete signal to audit key lengths. Legacy systems, embedded hardware, and older TLS configurations sometimes still use 1024-bit keys, a length that was flagged as insufficient years ago. The RSA-896 result adds urgency to those long-standing recommendations without changing the fundamental security analysis for properly configured modern systems. Two AI companies, Anthropic and Cognition, now have their models credited on consecutive public factoring records. Both Weis and Lu used AI models as collaborators on technically demanding work that previously required rare, highly specialized expertise. Claude helped Weis navigate the parameter space of a computation that would have taken a small team of number theorists months to configure manually.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our