Google DeepMind AI Agents Accelerate RSA-1024 Cryptoanalysis 1600-Fold
Google DeepMind agents identified vulnerabilities in the outdated RSA-1024 and accelerated the search for attacks on it by 1600 times. The study does not affect modern keys of 2048 bits or longer.
The results were announced by Pushmeet Kohli, Vice President and Senior Scientist at Google DeepMind. The research was conducted in collaboration with scientists from the University of California, San Diego, and the French center Inria.
AI agents examined the mechanism of digital signatures, in which fixed data is added before a message. According to the study, if an attacker controls approximately half of the block or more, such protective padding no longer enhances security: forging a signature becomes mathematically comparable to an attack on a system without additional data.
The proposed approach improved the result of the best-known attack on 1024-bit keys by 1600 times. It was tested in an isolated environment on a previously solved RSA-1024 benchmark. The authors also reported a reduction in costs for batch factorization—simultaneous analysis of multiple keys of the same length: computations for each subsequent key became 30 times cheaper compared to previous methods.
The authors emphasize that the findings do not threaten modern standards with key lengths of 2048 bits or longer, used in online banking, messengers, and digital currencies. In their assessment, automated cryptoanalysis could be used for stress-testing future standards and identifying risks before real threats emerge.
