Artificial intelligence has learned to crack RSA encryption 1,600 times faster

Researchers at Google DeepMind have used artificial intelligence to develop new cryptanalysis methods that allow for cracking outdated RSA encryption systems with 1,024-bit keys 1,600 times faster. The results of the study were published on November 6, 2026. The authors immediately emphasized that these methods do not affect modern systems with key lengths of 2,048 bits or more.

RSA is one of the best-known public-key encryption algorithms. Its security relies on the mathematical difficulty of factoring a large number into two prime factors. It is this principle that underlies digital signatures, which verify the authenticity of software, documents, and messages. The longer the key, the harder it is to break; therefore, 1,024-bit keys have long been considered obsolete and have been replaced by longer ones.

The research was conducted by a team led by Google DeepMind specialists DJ Dvidjotam and Abhradip Thakurta, together with cryptographers Nadja Henninger from the University of California, SanDiego and Emmanuel Tomé from the French institute Inria. The researchers used AI agents capable of independently developing mathematical approaches, writing code, conducting computational experiments, and correcting errors—all under human supervision.

The first line of research focused on the forgery of digital signatures. In some older systems, a special sequence of bytes—known as a padding—is added to the data before signing, which makes attacks more difficult. However, the researchers found that if an attacker controls approximately half or more of the block being signed, this protection effectively ceases to be an obstacle. In the case of a 32-byte digital fingerprint and a 1,024-bit key, AI helped identify an algorithm that theoretically speeds up the attack by approximately 1,600 times compared to the previous method.

This estimate is theoretical and not the result of a full-scale breach. In practice, the researchers tested the approach on a known 1,024-bit key: The AI agents optimized the software implementation of the algorithm and reduced the processor time required from 1,380 to 391 processor-years—a real-world speedup of about 3.5 times. Two AI agents completed all the work in less than four days: they wrote about 33,500 lines of code, conducted 336 distributed experiments, and utilized up to 300,000 processor cores and 16 graphics accelerators.

The second area of research concerns the mass factorization of RSA keys into prime factors. The new algorithm allows the results of previous computations to be reused immediately for a large number of keys of the same length. According to the authors' estimates, after the preparatory stage, the cost of processing each additional 1,024-bit key could be approximately 30 times lower than when using comparable previous methods. Some of the mathematical derivations have undergone formal verification in the Lean 4 system, but a full-fledged practical implementation has not yet been created.

The authors emphasize that their work pertains to specific scenarios involving outdated RSA, not to arbitrary attacks on modern secure connections. Nevertheless, the research demonstrates that AI is already capable of independently pushing the boundaries of cryptanalysis—and this is prompting security experts to monitor systems that still use short keys more closely.