Google DeepMind researchers, led by Dj Dvijotham and Abhradeep Thakurta, have identified two vulnerabilities in legacy 1024-bit RSA encryption. The findings, announced on November 6, 2026, were achieved through human-guided AI models that accelerated the exploration of mathematical assumptions.
The first identified weakness allows for the forgery of digital signatures, while the second involves a new algorithm that reduces the computational cost of batch factoring for related moduli. These vulnerabilities suggest that attacking 1024-bit RSA systems requires significantly less effort than previously estimated by cryptographers.
The research team collaborated with experts Nadia Heninger and Emmanuel Thomé to demonstrate how AI can perform exploratory tasks faster than traditional human-led efforts. The study notes that modern RSA implementations using 2048-bit keys or larger remain secure and are not affected by these specific threats.
This development coincides with other 2026 findings regarding potential attacks on post-quantum cryptographic candidates, such as HAWK. The research highlights a shift in the speed at which vulnerabilities in encryption algorithms can be uncovered when AI agents are utilized to assist human researchers.
