AI agent factors RSA-260, setting a new record and pegging RSA-1024 at ~$30M

Cognition researcher used the company’s autonomous coding agent, Devin, to factor RSA-260, the largest number ever cracked in the RSA Factoring Challenge and the first new record since RSA-250 fell in February 2020. Rather than any mathematical breakthrough, the feat came from performance engineering: Devin built a heavily modified version of CADO-NFS that runs the general number field sieve — specifically the lattice-sieving stage — on GPUs, reportedly cutting the cost of large factorizations by roughly 10x versus the prior public state of the art. The whole job consumed about 4,900 GPU-days (~$400k at market rates) and ran opportunistically on idle, fragmented nodes left over from the company’s LLM training clusters, since lattice sieving is embarrassingly parallel and safe to preempt.

The practical takeaway is about who can now do this and for how much. The author estimates that a hyperscaler or frontier AI lab could factor a 1024-bit RSA key for around $30 million — a figure that further optimization could plausibly halve. That RSA-1024 is breakable isn’t news (deprecated since 2013, with NSA capability speculated since the mid-2000s), but the barrier has shifted: you now need only a large pool of commodity GPUs rather than custom hardware like TWIRL, and non-cryptographers can meaningfully contribute. RSA-2048 remains about a billion times harder and is essentially untouched by this work.

The more provocative claim is about the agent itself. The author frames their own role as setting priorities, defining benchmarks, and catching dead ends, while Devin autonomously handled measurements, cluster operations, and end-to-end optimization at the intersection of computational number theory and GPU kernel engineering — work that would normally demand a multi-month effort from specialized experts. The broader argument: autonomous software agents are sharply lowering the barrier to entry for cryptanalysis and large-scale scientific computing generally.