DeepMind says its AI agents ran a four-day cryptography test with about 20 human messages
Google DeepMind reports that two AI agents reproduced a previously solved 1024-bit RSA-signature computation in three days and 22 hours, using 391 CPU core-years—about 3.5 times less than an earlier run’s reported CPU use. The agents wrote and tested code, ran distributed jobs, debugged an overflow error and verified the result, with about 20 human messages largely approving decisions. The task involved billions of advance queries to a simulated signing service; the agents did not recover a private key or make further queries after access ended. The work used a sandboxed benchmark and does not threaten deployed RSA systems at 2048 bits or larger. The comparison is not controlled, and GPU work and preliminary runs were excluded. A separate proposed algorithm remains unimplemented.
