Google DeepMind says two AI agents completed a known 1,024-bit RSA-related computation in 3 days and 22 hours, handling code development, testing, distributed jobs and troubleshooting with about 20 researcher messages, mostly approvals. The experiment shows agents can carry out complex computing tasks within human-defined limits, but it does not validate a new theoretical algorithm or indicate an immediate threat to widely deployed RSA systems.
A known RSA signature calculation
The test focused on RSA digital signatures. A simulated signing service accepted about 4 billion advance queries before the target was selected. The service then stopped accepting queries, and the agents calculated a mathematical root needed to produce a new signature without making further queries. The work did not recover the private key.
DeepMind researcher Pushmeet described the work on November 6. The paper characterizes it as a sandboxed reproduction of a benchmark instance that had already been solved, and says the result does not threaten deployed cryptographic systems using RSA keys of 2,048 bits or longer.
Lower CPU use, but not a like-for-like comparison
The paper reports that the computation used 391 CPU core-years, compared with about 1,380 core-years in an earlier run by a team using the same instance and algorithm. Based on the reported figures, that is roughly a 3.5-fold reduction. The runs used different hardware, however, so the comparison is not strictly like-for-like; the figure also does not represent total financial cost.
The 391 core-years exclude GPU work and about 7.5 core-years spent on preliminary probing and rehearsals. The experiment also set aside about 330 hours of GPU time, of which roughly 100 hours were active. DeepMind says about 3.2 days elapsed between the start of the first full-scale sieving task and verification of the result. At peak, the run used about 300,000 CPU cores and 16 GPUs. Researchers attributed the difference from an earlier run that took about five months of calendar time mainly to parallel computing capacity.
Researchers set limits as agents split the work
Researchers chose the problem and specific instance, set limits on the signing service, and approved conditions before irreversible steps. The paper says they did not write code, launch computing jobs or diagnose failures.
An execution agent wrote and tested programs, launched distributed jobs and fixed operational problems. A supervisory agent tracked the design, resource budget and checks, and used a specialized sub-agent to accelerate the linear algebra stage. Together, the two agents added about 33,500 lines of code, ran 336 distributed experiments and exchanged 218 messages.
The first mathematical root was wrong because a 32-bit offset overflowed. Correctness checks caught the error, and the agents identified and fixed it without repeating the advance signature queries. The process took 3 days and 22 hours from design through verification. The roughly 20 human messages and approval steps also show that the run was not fully unsupervised.
The result does not validate a new algorithm
The reported reduction in CPU use came from improvements to the implementation and execution of an existing algorithm. The experiment did not demonstrate a new algorithm for more restricted signature-access conditions.
A separate paper on “Fan Factories” estimates that, after shared precomputation costs roughly equivalent to one general number field sieve factorization, the cost of each additional 1,024-bit target could be about 30 times lower than with the multiple number field sieve. That remains an asymptotic estimate; no Fan Factory has been implemented to date.
The working paper does not disclose the specific names of the main models, referring to them only as A and B, and does not link to a public implementation repository. That limits outside researchers’ ability to compare the models independently. The reported CPU figure, GPU resources and human approval requirements are all relevant constraints when assessing the result.