Two artificial intelligence agents from Google DeepMind reproduced a known 1024-bit RSA cryptography computation in a run lasting just under four days, during which the human researcher overseeing the effort sent about 20 messages — most of them simple approvals — and wrote no code, according to a working paper from the company.

Over the course of the run, the agents produced roughly 33,500 lines of code, exchanged 218 messages and executed 336 distributed experiments, completing the task with about 3.5 times less CPU use than an earlier effort on the same instance. The papers state that the results do not threaten deployed systems that rely on 2048-bit or larger RSA keys.

The assignment was a sandboxed reproduction of a benchmark instance tied to RSA digital signatures — a cryptographic problem that another team of researchers had already solved. A Google DeepMind researcher picked the problem, set the rules and then largely stepped aside. Once the run was underway, the two agents handled the coding, the computing and the debugging on their own, and the paper says the researcher wrote no code, launched no jobs and diagnosed no faults.

DeepMind researcher Pushmeet Kohli announced the work on Nov. 6.

— Pushmeet Kohli (@pushmeet) November 6, 2026

How the Work Was Split

The researcher chose the problem, the instance and the limits placed on a simulated signing service, and approved the conditions for any step that could not be undone. A supervisor agent tracked the design, budgets and checks, while an executor agent wrote and tested code, launched distributed jobs and fixed failures. The supervisor also brought in a specialist sub-agent to speed up one linear algebra step.

From design to the verified result, the work took three days and 22 hours.

The setup unfolded in several stages. A separate program that held the modulus factorization answered roughly 4 billion advance queries, then was withdrawn before the target was chosen. The agents then computed the mathematical root needed for a new signature without issuing any further queries. They did not recover a private key at any point.

A Bug Caught Along the Way

The first root the agents computed was wrong. The cause was a 32-bit offset that overflowed, and a correctness check flagged the error. The agents located and repaired the fault without repeating the advance signing queries.

According to the paper, the completed computation used 391 CPU core-years, while an earlier team needed a rounded 1,380 core-years on the same instance and algorithm — roughly 3.5 times less CPU use for the DeepMind agents. The paper notes that this figure excludes about 330 reserved GPU-hours, of which about 100 were busy, as well as 7.5 core-years of probes and rehearsals. The two totals also come from different hardware, so the comparison is not like-for-like.

The earlier academic effort took about five months. DeepMind's run needed 3.2 days from its first full-scale sieving job to the verified root, peaking at about 300,000 CPU cores and 16 GPUs, and the authors credit parallelism for most of that gap.

The run did not test the paper's new algorithm for more restricted signing access at full cryptographic size. A separate Fan Factories paper estimates about 30-fold lower costs per additional 1024-bit target after shared precomputation, but no fan factory has been built.

The working paper identifies its main models only as A and B and does not link to a public code repository, which limits outside comparison. Independent replication and testing beyond the 1024-bit benchmark therefore remain open points for assessing how far the result generalizes.