Google DeepMind: AI agents found two weaknesses in legacy 1024-bit RSA

Google DeepMind: AI agents found two weaknesses in legacy 1024-bit RSA

📅 November 7, 2026 • Google DeepMind

Two working papers from Google DeepMind
Google DeepMind announced on November 6, 2026 that AI agents working under human supervision had found two weaknesses in legacy RSA cryptography on 1024-bit keys. The results are described in two working papers published on November 5 and November 6, 2026, Modular Roots from a Restricted Oracle and Fan Factories. Co-authors include Krishnamurthy (Dj) Dvijotham, Abhradeep Thakurta, Nadia Heninger of UC San Diego and Inria, Emmanuel Thomé of Inria, Pushmeet Kohli and others. Neither paper has been peer reviewed.

Result one: forging digital signatures
The first result concerns digital signature forgery. Per the papers, a theoretical estimate gives about 1600 times a better result than the previous best method for 1024-bit keys, in the case where the attacker controls a small part of the signed data, for example a 32-byte digital fingerprint. The improvement is stated as a theoretical estimate, and the work was carried out in an isolated environment.

Result two: batch factorization
The second result is about batch factorization. Every integer gets its own family of number fields, so under a budget for one factorization the next 1024-bit number is about 30 times cheaper to factor than with the multiple number field sieve method (MNFS). In cost terms that is roughly 1/58 of what factoring with a general number field sieve costs.

Reproducing an earlier 1024-bit computation
The agents also reproduced an earlier 1024-bit computation. The first root took 391 CPU core-years, which is 3.5 times less than the original computation. The work ran for 3.2 days from the start of full-scale sieving, on up to 300,000 CPU cores and 16 GPUs, against about five months for the previous computation. One agent generated the mathematical proofs and a second agent verified them in Lean 4.

Scope and what it means for engineers
The agents found and resolved four listed problems that affected the algorithm or the correctness of the result, and none of them required another oracle pass. Google DeepMind states directly that modern systems with keys of 2048 bits and above are not affected, and that the results were obtained in an isolated environment and do not touch running systems. For engineers, cryptographic hardness can no longer be treated as a fixed assumption, and the research cycle with AI agents is now visible in a field that usually moves slowly.

IMPACT An agent runtime that proves and checks its own math turns a question of key sizes into a question of research speed, and that is a planning input for anyone who assumes cryptography is a solved problem.

Sources: storage.googleapis.com - interestingengineering.com - aichatdaily.com

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky