DeepMind announced that two of its AI agents successfully reproduced a known 1024‑bit RSA digital‑signature benchmark in just under four days, using roughly 391 CPU core‑years of compute. The effort, described in a working paper released on November 10, 2026, involved the agents handling coding, large‑scale execution and debugging, while human operators intervened only to approve key steps.

Process and human oversight

According to CryptoSlate, the workflow was split between a supervisor agent, which managed design, budgets and safety checks, and an executor agent that wrote and tested code, launched distributed jobs and repaired failures. A specialist sub‑agent was also called upon to speed up a linear‑algebra sub‑task. Human operators selected the problem instance, set limits on a simulated signing service, and approved irreversible actions, resulting in roughly 20 human messages throughout the run.

Compute efficiency and scale

The agents ran 336 distributed experiments, generated about 33,500 new lines of code and exchanged 218 internal messages. At peak, the system employed around 300,000 CPU cores and 16 GPUs, though GPU usage and preliminary probes are excluded from the reported 391 CPU core‑year figure. Compared with an earlier academic effort that required about five months and 1,380 core‑years, DeepMind’s run achieved a 3.5‑fold reduction in CPU consumption, primarily due to implementation and execution optimizations rather than algorithmic breakthroughs.

Cryptographic implications

The test reproduced a known RSA root without recovering the private key and did not demonstrate the new theoretical algorithms presented in the associated research papers. The authors stress that the results do not threaten real‑world RSA deployments using 2048‑bit keys or larger. The benchmark involved a simulated signing service that answered roughly four billion advance queries before being withdrawn, after which the agents computed the target root without further queries.

What to watch next

Future work will focus on extending the agents’ capabilities to larger key sizes and evaluating the practical security impact of AI‑driven cryptanalysis. Observers will also monitor whether the reported CPU savings can be replicated on comparable hardware and whether the approach can be adapted to other cryptographic challenges.