Cognition's Devin Agent Breaks RSA-260 Factoring Record

Cognition researcher Eric Lu used the company's Devin AI agent to factor RSA-260, establishing a new record for the RSA Factoring Challenge by surpassing 2020's RSA-250 benchmark. Over a three-week side project, Lu directed up to 18 concurrent Devin sessions to rewrite CADO-NFS, a standard open-source CPU-based codebase, to run on NVIDIA GB200, GB300, and B200 GPUs. The resulting GPU-optimized implementation, which included a new lattice siever named glas, ran on idle single-node fragments within NVL72 racks left over from large language model training.

The entire factorization process consumed approximately 4,900 GPU-days, translating to roughly $400,000 at current market rates. This compute budget included 643 GPU-days for polynomial selection and 467 GPU-days for solving linear systems. The achievement represents a tenfold cost reduction compared to previous state-of-the-art methods. Based on these metrics, factoring the deprecated RSA-1024 standard would require 78 times more computation, costing an estimated $30 million, while the highly secure RSA-2048 standard remains mathematically out of reach.

Devin operated across 192 sessions, consuming 14,450 ACUs and exchanging 82,702 words over 3,328 messages. Notably, 36 of these sessions required no human intervention. The agent modified nearly every pipeline stage, adapting polynomial selectors, optimizing block Wiedemann implementations, and rewriting a GPU-accelerated square root function three times. Lu, who does not specialize in number theory, acted in an executive capacity by setting goals, correcting errors, and suggesting optimization paths like using NVLink SHARP.

For software engineers and researchers, this milestone demonstrates that autonomous agents can successfully execute high-level performance engineering. By substituting for a multi-month effort by domain experts, Devin shows that a non-expert can now tackle specialized mathematical problems. However, the experiment also revealed limitations: the agent performed best when working with recognizable reference code and grew increasingly confused as the codebase diverged from its pretraining data.