Google DeepMind Repeats RSA-1024 Benchmark With 391 CPU Core-Years

Two primary AI agents and a specialist sub-agent at Google DeepMind reproduced a solved RSA-1024 computation with 391 CPU core-years, about 3.5 times below an earlier run on the same instance and algorithm.

The task was a sandboxed benchmark involving RSA digital signatures. The agents computed the mathematical root needed for a signature; they did not recover a private key. The project took three days and 22 hours from design through verification. Measured from the first full-scale sieving job to the verified root, the run lasted 3.2 days.

Before the target was chosen, a separate program holding the RSA modulus’s factorization answered roughly 4 billion advance queries for a simulated signing service. The program was then withdrawn, and the agents computed the root without making further queries.

A human researcher chose the problem and instance, set limits for the simulated signing service and approved steps that could not be undone. The researcher sent about 20 messages, mostly approvals, and wrote no code. The agents produced about 33,500 lines of code and ran 336 distributed experiments.

The 391 CPU core-years exclude about 330 reserved GPU-hours, of which roughly 100 were busy, and 7.5 core-years spent on probes and rehearsals. The run peaked at about 300,000 CPU cores and 16 GPUs. The earlier effort took about five months, but it used different hardware, so the CPU totals are not a like-for-like performance comparison.

The 3.2-day measurement covers the interval from the first full-scale sieving job to the verified root. Parallelism accounts for most of the calendar-time gap with the earlier effort. The run did not test the research’s separate algorithm for more restricted signing access at full cryptographic size.

The findings do not threaten deployed systems using RSA keys of 2,048 bits or larger. The result follows Google DeepMind’s earlier report on two RSA-1024 weaknesses.