Google DeepMind: AI agents found two weaknesses in legacy 1024-bit RSA

On November 7, 2026, Google DeepMind announced that AI agents working under human supervision had discovered two weaknesses in legacy RSA cryptography for 1024-bit keys. The findings were detailed in two working papers, "Modular Roots from a Restricted Oracle" and "Fan Factories," published on November 5 and November 6, respectively. These papers have not yet undergone peer review.

The first weakness identified pertains to digital signature forgery. Theoretical estimates suggest that AI agents can forge digital signatures approximately 1600 times more effectively than the previous best method for 1024-bit keys, given that the attacker controls a small portion of the signed data. For example, if a digital fingerprint consists of 32 bytes, the AI agents' method is theoretically 1600 times more efficient.

The second weakness relates to batch factorization. In this scenario, each integer is associated with a distinct family of number fields. The AI agents demonstrated that factoring the subsequent 1024-bit integer is roughly 30 times cheaper compared to the traditional multiple number field sieve (MNFS) method. In practical terms, this translates to about 1/58 of the computational cost when employing the general number field sieve.

Additionally, the AI agents reproduced an earlier 1024-bit computation. The first root took 391 CPU core-years, which is 3.5 times less than the original computation. The entire process spanned 3.2 days, from the initiation of full-scale sieving to the conclusion, involving 300,000 CPU cores and 16 GPUs. In contrast, the previous computation required about five months to complete.

One agent was responsible for generating mathematical proofs, while a second agent verified these proofs using Lean 4. The agents identified and addressed four issues affecting the algorithm's correctness or its results, with no additional oracle passes required. Google DeepMind emphasized that modern systems employing keys of 2048 bits and above remain unaffected by these findings. Furthermore, the research was conducted in an isolated environment, and the AI agents did not interfere with running systems.

Engineers should no longer assume cryptography is a fixed and unchanging problem. The research cycle involving AI agents is now visible in a field that typically progresses slowly. With AI agents capable of proving and checking their own mathematical work, the question of key size selection has shifted from a static assumption to a matter of research speed. This development presents a planning input for anyone who continues to rely on cryptography as a solved problem.