DeepMind: AI agents accelerate cryptanalysis, finding two weaknesses in legacy RSA-1024
Google DeepMind said human-guided AI agents found two major weaknesses in legacy 1024-bit RSA: limits of digital-signature padding protection, and a new batch key-factoring algorithm. It stresses that modern 2048-bit and larger encryption is not affected by these specific results.
Google DeepMind announced on November 7, 2026 (Singapore time) new research showing that AI agents, guided by humans, can accelerate cryptanalysis—the science of stress-testing encryption—and uncovered two major weaknesses in legacy 1024-bit RSA cryptosystems.
According to an official essay by DeepMind VP of Science Pushmeet Kohli, the results apply only to obsolete 1024-bit RSA. Modern encryption, which almost universally uses 2048-bit or larger keys, remains secure against these specific findings. DeepMind says it is sharing the work so the security community can better measure cryptographic strength and keep standards resilient ahead of threats.
The first result concerns digital signatures. Many signature standards add “padding” before signing. The team found that what matters is how much of the signed block an attacker controls: once they control roughly half or more, padding adds no protection. In a sandboxed simulation that did not touch live systems, agents reproduced a result on an already-solved 1024-bit benchmark key from earlier academic work with about 3.5× less CPU time, and at less than a thousandth of the computing power needed to break the key itself. When an attacker controls less of the block—for example a short 32-byte fingerprint—forgery is harder; even then, DeepMind says a new algorithm improves on the prior best method by roughly 1,600× for 1024-bit keys on theoretical estimates.
The second result concerns factoring many same-size moduli. Agents found a new algorithm with better asymptotic complexity than prior batch methods, cutting the average cost of factoring each additional 1024-bit key among many by about 30× versus previous batch approaches.
The work was led by Dj Dvijotham and Abhradeep Thakurta, in collaboration with Nadia Heninger (UC San Diego and Inria) and Emmanuel Thomé (Inria). Technical papers on digital signatures and key factorization were linked from the announcement.
Official lab post: Google DeepMind; detailed essay: Advancing the frontier of cryptanalysis