Google DeepMind: AI has made old RSA cheaper to attack, but has not broken modern encryption

When you log in to online banking or open a secure website, the systems use encryption to keep data secret. One of the best-known methods is called RSA. It relies on the fact that it is very hard to split a large number into its prime factors. Now Google DeepMind says AI agents have found two weaknesses in an older version of RSA.

They apply only to RSA with 1024-bit keys. According to Crypto Briefing, the results do not apply to RSA with 2048 bits or more and do not affect modern encryption solutions. The work was also carried out in an isolated environment and does not touch systems in operation.

Two weaknesses

The first weakness concerns digital signatures, which are used to prove who a message comes from. DeepMind has found a method for forging them in older 1024-bit RSA. According to theoretical estimates, the new algorithm improves on the best previous method by roughly 1,600 times.

The agents tested the method on a 1024-bit key that researchers had already solved in an earlier academic study. It produced the same result using roughly 3.5 times less CPU time. The earlier study took roughly five months. For the first root, 391 CPU-core years are cited, meaning computing time equal to one processor core running for 391 years.

The second weakness is a new algorithm that can break many 1024-bit keys together. According to DeepMind, it lowers the average cost of each additional key in a large collection by roughly 30 times compared with earlier batch methods.

What is behind it

The results are described in two working papers, published on November 5 and 6, 2026: "Modular Roots from a Restricted Oracle" and "Fan Factories." The work was done together with cryptanalysis experts, including Nadia Heninger of UC San Diego and Inria and Emmanuel Thomé of Inria.

Humans guided the exploration, while the AI models did the exploratory work. According to Crypto Briefing, the main finding in DeepMind's announcement is that AI agents can sharply shorten the time it takes to examine the mathematical assumptions that encryption algorithms rely on.

Caveats

The figures come from DeepMind's own claims, as reported by other outlets. According to the sources, the two papers have not been peer reviewed, meaning they have not been vetted by independent researchers. The 1,600-fold figure is a theoretical estimate, and it is only stated that the method was tested on an already solved key, not on keys in use.