Research New
Google DeepMind says human-guided AI agents found two weaknesses in legacy 1024-bit RSA
Published:
Google DeepMind announced on the evening of 6 November, Paris time, two working papers written with cryptographers Nadia Heninger and Emmanuel Thomé in which AI agents, supervised by researchers, attacked legacy 1024-bit RSA. Both papers state that deployed RSA keys of 2048 bits or more are not threatened.
The first paper shows that the PKCS#1 v1.5 signing format stops adding protection once an attacker may submit more than about half a block of data, or a 64-byte digest at 1024 bits. It also reports a re-run of a published 1024-bit signature computation: the first root took 391 CPU core-years, 3.5 times less than the original, and two cooperating AI agents designed, ran and debugged it in under four days.
The second paper, dated 5 November, is theoretical. With a precomputation costing about one factorisation, each further 1024-bit key could be factored at about 1/58 of the usual cost, 30 times less than with the best earlier method for that budget. One model generated the proofs and a second checked them, with the algebra machine-checked in Lean 4.
Caveat. Both are working papers, not peer reviewed. The factoring gain is a leading-order estimate with lower-order terms set to zero; no factorisation factory was built and no keys were factored.
Sources