Cognition AI Factors RSA-260 Using AI-Driven GPU Optimization

Eric Lu and a research team at Cognition AI successfully factored RSA-260 on September 3, 2026, marking a significant milestone in computational number theory. The achievement represents the largest publicly solved problem from the RSA Factoring Challenge to date, surpassing the previous record of RSA-250.

To solve the 862-bit challenge number, the team implemented a high-performance version of the General Number Field Sieve (GNFS) specifically optimized for GPUs. According to Cognition AI, the process was driven by a fleet of "Devins"—the company's AI agents—which were used to build and optimize the GPU lattice siever. Eric Lu claimed that this AI-driven approach enabled the factorization to be completed at a cost 10 times lower than the previous public state of the art, though this specific efficiency gain has not been independently verified by a third-party audit.

The RSA Challenge

RSA-260 is part of a series of challenge numbers created by RSA Laboratories in 1991. These numbers were designed to test the limits of integer factorization, the mathematical process of finding the two large prime factors of a semi-prime modulus. Because the security of RSA encryption relies on the extreme difficulty of this task, these challenges serve as critical benchmarks for the cryptographic community to gauge the evolution of compute power and algorithmic efficiency.

Implications for Research

While the factorization of RSA-260 does not mean that modern RSA encryption is broken—current industry standards typically use much larger keys, such as RSA-2048—the methodology used by Cognition AI signals a shift in scientific research. The use of AI agents to optimize low-level hardware performance and complex mathematical sieving suggests that AI can now accelerate the development of highly specialized computational tools that previously required manual, expert-level engineering.

What's Next

The success of the RSA-260 project highlights the increasing viability of GPU-accelerated number theory. Future attention will likely focus on whether this AI-driven optimization can be scaled to even larger challenge numbers or if the 10x cost reduction claimed by Cognition AI can be replicated across other cryptographic benchmarks. For now, the result stands as a demonstration of how autonomous agents can be applied to solve long-standing mathematical hurdles.