DeepMind AI agents find two weaknesses in legacy 1024-bit RSA cryptosystems

Google DeepMind researchers used human-supervised AI agents to identify weaknesses in legacy 1024-bit RSA systems, arguing that AI can accelerate the testing of cryptographic assumptions that defenders have traditionally expected to change slowly. The work focuses on two areas: digital-signature forgery and factoring multiple RSA keys of the same size. The researchers say the findings apply to obsolete 1024-bit systems, while modern protocols generally use keys of 2048 bits or more and remain secure.

For signatures, the team found that padding may offer no protection when an attacker can control roughly half or more of the data block being signed. In a sandboxed test using a previously solved 1024-bit benchmark key, agents reproduced an attack using 3.5 times less CPU time than the earlier approach, without testing live systems. For cases where the attacker controls less data, the researchers estimate a new method could improve on the prior best technique by about 1,600 times for 1024-bit keys. Separately, they developed a method for factoring many same-sized RSA numbers, reducing the average cost of each additional 1024-bit key by about 30 times relative to earlier batch methods. The broader warning is that cryptographic strength should be continually reassessed as AI speeds mathematical research.