Google DeepMind says two AI agents carried a 1024-bit RSA cryptography computation through to a verified result in just under four days, with roughly 20 human messages along the way. Most of those messages were approvals rather than instructions. The humans picked the problem instance and set the limits. The agents did the rest.
What the agents actually did
The setup is a sandboxed reproduction of a benchmark instance that researchers had already solved before. A supervisor agent held the design, budgets and checks. An executor wrote and tested code, launched distributed jobs and repaired failures. The supervisor also brought in a specialist sub-agent for one linear-algebra step.
Over the run the agents produced about 33,500 new lines of code, ran 336 distributed experiments and exchanged 218 messages between themselves. From design to verified result took three days and 22 hours.
The work involved billions of advance queries to a simulated signing service. Once that access was withdrawn, the agents computed the mathematical root needed for a new signature. They did not recover a private key. That distinction matters.
Where the savings came from
The paper reports 391 CPU core-years for the finished computation. An earlier team, using the same instance and algorithm, reported a rounded 1,380. That is roughly a 3.5-fold reduction in measured CPU use, though GPU work and early probes are excluded from the newer figure.
The earlier academic run took about five months. DeepMind’s took 3.2 days from its first full-scale sieving job to the verified root, peaking near 300,000 CPU cores and 16 GPUs. The authors put the calendar-time gap down mainly to parallelism. The CPU saving, they say, came from implementation and execution changes inside the existing algorithm.
One hiccup: the first computed root was wrong because a 32-bit offset overflowed. A correctness check caught it, and the agents found and fixed the fault without redoing the advance signing queries.
What this does not show
DeepMind researcher Pushmeet announced the work on Nov. 6. The paper states plainly that the results do not threaten deployed systems using 2048-bit or larger RSA.
Also worth noting, the CPU totals come from different hardware, so this is not a controlled like-for-like comparison. And the measured reproduction did not exercise the paper’s newer algorithm for tighter signing access at cryptographic size. That side rests on a separate heuristic estimate in a Fan Factories paper, and no fan factory has been built.
The models behind the run appear only as A and B, and no implementation repository is linked. That limits outside comparisons, perhaps for now.
