Google DeepMind says two AI agents reproduced a known 1024-bit cryptography computation, handling programming, large-scale execution and debugging with about 20 human messages, mainly approvals. The agents reached a verified result three days and 22 hours after design work began. A DeepMind researcher announced the work on November 6 as a sandboxed reproduction of a benchmark that had already been solved.
The experiment involved RSA digital signatures rather than recovering a private key. A separate program with the modulus factorization supplied answers to roughly 4 billion preliminary queries. Access to that simulated signing service ended before the target was selected. The agents subsequently calculated the mathematical root required for a new signature without making further queries. The papers say the results pose no threat to deployed cryptosystems using RSA with keys of 2048 bits or more.
The human researcher selected the problem, the instance and the restrictions on signing-service access, then approved conditions governing irreversible steps. The paper says the researcher neither wrote code nor started jobs nor diagnosed failures. A supervisor agent managed the design, resource budgets and verification checks. An executor agent developed and tested code, started distributed jobs and corrected failures. The supervisor recruited a specialist sub-agent to speed up a linear-algebra step. The agents generated about 33,500 lines of new code, conducted 336 distributed experiments and sent 218 messages to each other.
A 32-bit offset overflow caused the agents' initial root calculation to be incorrect. A verification check detected the problem, and the agents identified and repaired the fault without repeating the preliminary signing queries. The small number of human messages therefore did not remove human task selection or approval requirements.
The working paper puts CPU use for the completed calculation at 391 core-years, against an earlier team's rounded total of 1,380 for the same instance and algorithm. The figures represent approximately a 3.5-fold reduction in measured CPU use. The earlier academic effort lasted about five months. DeepMind's computation reached the verified root 3.2 days after the first full-scale sieving job began. Peak resources reached about 300,000 CPU cores and 16 GPUs. The authors attribute the shorter elapsed time chiefly to parallel execution.
Implementation and execution changes to the existing algorithm produced the CPU savings. The 391-core-year figure omits about 330 reserved GPU-hours, including about 100 hours of active GPU use. The total also omits 7.5 CPU core-years spent on probes and rehearsals. Differences between the hardware used for the two CPU totals prevent a controlled comparison under identical conditions.
The measured reproduction did not demonstrate the working paper's new algorithm at cryptographic scale under more restrictive signing access. A separate paper, Fan Factories, estimates that each additional 1024-bit target would cost roughly 30 times less than the multiple number field sieve. That estimate assumes shared precomputation costing approximately the equivalent of one general number field sieve factorization. The cost projections are leading-order heuristic estimates, and no fan factory has been implemented. The working paper labels the principal models only A and B and provides no link to a public implementation repository, restricting outside comparisons.