Google DeepMind AI agents identify weaknesses in legacy RSA-1024 cryptography

Google DeepMind has announced that AI agents, operating under human supervision, have discovered two significant weaknesses in legacy 1024-bit RSA cryptography. The findings, detailed in two non-peer-reviewed working papers, involve digital signature forgery and batch factorization.

The first discovery relates to digital signature forgery. Researchers estimate that if an attacker controls a small portion of signed data, such as a 32-byte digital fingerprint, the new method is approximately 1,600 times more effective than previous best methods for 1024-bit keys.

The second finding concerns batch factorization. The research suggests that once the initial cost of one factorization is paid, subsequent 1024-bit numbers are roughly 30 times cheaper to factor than with the multiple number field sieve method. In terms of total cost, this is approximately 1/58th of the cost of factoring with a general number field sieve.

In testing, the AI agents reproduced an earlier 1024-bit computation using 391 CPU core-years, which is 3.5 times less than the original computation. The process took 3.2 days using up to 300,000 CPU cores and 16 GPUs. During the process, one agent generated mathematical proofs while a second agent verified them using Lean 4.