AI agents guided by researchers find two cheaper routes to attack outdated 1024-bit RSA keys

Google DeepMind on Friday presented two working papers in which AI agents, guided by researchers, found two possible weaknesses in RSA encryption with 1024-bit keys. DeepMind's Pushmeet Kohli announced the work. According to DEV Community, the co-authors include cryptographers Nadia Heninger (UC San Diego and Inria) and Emmanuel Thomé (Inria). According to Transclew, neither method has yet been carried out at full scale.

The first weakness concerns forging signatures when an attacker can supply part of the data to be signed to a signing device. That requires direct access and billions of requests. DeepMind speaks of a theoretical improvement of roughly 1,600 times, but according to Transclew that figure appears only in the announcement; the paper itself describes it as a change in a cost formula. The second method makes cracking several keys together cheaper: after an up-front investment, each further key would cost about thirty times less than with the best previous method.

What was actually measured is that the agents repeated a known forgery on a test key using 391 processor core-years, about 3.5 times less than in an earlier study, which is attributed mainly to a better implementation. The US standards body NIST no longer accepts 1024-bit keys for new signatures, but the DKIM email standard still allows them. By Saturday, no other group had reproduced the results.