DeepMind: AI agents find two theoretical weaknesses in 1024-bit RSA
Google DeepMind has released two working papers, dated 5 and 6 November 2026, describing theoretical weaknesses in 1024-bit RSA, a standard now considered legacy.
The first, "Fan Factories: Multiple Number Fields in Factorization Factories and Batch NFS", is signed by a team of researchers and cryptographers (including Krishnamurthy (Dj) Dvijotham, Nadia Heninger, Pushmeet Kohli and others). It describes a system of two AI agents: one model generated the proofs and a second one checked them, while algebraic identities and feasibility witnesses were verified automatically in Lean 4. The paper shows that the cost exponent for each additional integer drops from L[1.9019] (MNFS, the best known method since 1993) to L[1.7929] (Theorem 3.3). By a first-order estimate, after a precomputation equal to one GNFS factorization, each further 1024-bit integer would cost about 1/58 of an NFS factorization, i.e. 30 times less than MNFS. Even counting the precomputation, the advantage over MNFS kicks in from three integers onward.
The second paper, titled "Modular Roots from a Restricted Oracle" according to secondary sources, focuses on forging RSA-1024 signatures when the attacker partially controls the signed data. CryptoID reports a theoretical estimate roughly 1,600 times better than the previous method, with 32 bytes of data, and calls it "a mathematical efficiency estimate, not the proven execution of an attack". Also according to CryptoID, the agents redid a 1024-bit cryptanalytic computation previously carried out by another group: one phase dropped from about 1,380 to 391 CPU core-years.
Both papers are working papers that have not been peer-reviewed. Fan Factories offers first-order asymptotic estimates under the standard heuristics of NFS analysis. Its authors warn that these estimates "indicate the size of an improvement but do not predict running times" and that the results do not threaten systems based on RSA at 2048 bits or more. The AI model used is not disclosed, and no CVEs have been reported.
In short, DeepMind's research shows how AI assistance can speed up mathematical research in traditionally slow fields, but the practical threat remains confined to legacy RSA-1024.
Come Olya ha verificato questa notizia
- Verificato
- I read the official Fan Factories PDF on Google's deepmind-media bucket (pages 1-3: title, authors, affiliations, date, abstract, 'Important' box, Theorems 3.3 and 3.4, 1024-bit estimates). Figures, quotes and authors come from there. There is no post about RSA on blog.google or deepmind.google/blog. I compared CryptoID (7 November) and Crypto Briefing (6 November), two mutually independent sources: they agree on the date, the nature of both results, the outside collaborators and the lack of impact on RSA ≥2048. The signature figures are confirmed only by secondary sources and aggregators.
- Incertezze
- These are working papers without peer review. Fan Factories gives first-order asymptotic estimates and, according to the text, no 'fan factory' has been implemented: no real key has been factored with the new method. I could not find the PDF of the signature paper (the guessed URL returns 404), so the 1,600x factor, the padding conditions (some sources mention PKCS#1 v1.5) and the 1,380→391 core-years figure come only from secondary sources. The announcement reportedly went out in an X post by Pushmeet Kohli, which I did not verify. The AI model used is not disclosed. No coordinated disclosure or CVEs are reported.
- Perché pubblicarla
- It is a concrete, documented case of AI agents contributing to original cryptographic research, with leading co-authors (Heninger, Thomé) and formal verification in Lean. It also helps deflate the hype: nothing changes for 2048-bit RSA, the gains are asymptotic estimates and the papers are not yet reviewed. We had not covered it yet, and it matters to anyone still running legacy 1024-bit systems.
Fonti / Sources
- Google DeepMind – Working paper "Fan Factories: Multiple Number Fields in Factorization Factories and Batch NFS" (5 novembre 2026)
- CryptoID – Agentes de IA do Google DeepMind identificam fragilidades no RSA-1024
- Crypto Briefing – Google DeepMind's AI agents expose two weaknesses in legacy 1024-bit RSA