Digital Tech News Network banner
DTN News

Artificial Intelligence Agents Found Two Weaknesses in Retired RSA Encryption

Abstract illustration of a golden key made of circuit lines dissolving into dots on a dark navy background

Google DeepMind says its artificial intelligence agents, working under human supervision, found two new weaknesses in retired 1024-bit RSA encryption. The headline numbers: roughly 1,600 times better at forging signatures, 30 times cheaper at factoring batches of keys. But both of those numbers are theoretical estimates, not things anyone ran. What the agents actually ran is smaller — and more interesting.

Google DeepMind logo
Google DeepMind — the AI lab behind Gemini.

What happened

On November 6, 2026, Google DeepMind's Pushmeet Kohli published "Advancing the frontier of cryptanalysis," describing two results from AI agents working under human supervision. The team was led at DeepMind by Krishnamurthy (Dj) Dvijotham and Abhradeep Thakurta, with cryptographers Nadia Heninger and Emmanuel Thomé. The work is written up in two working papers, and both carry the same warning in bold: the results do not threaten cryptosystems based on 2048-bit or larger RSA.

Abstract illustration from Google DeepMind's announcement essay: a white block split open to reveal a maze of translucent glass walls
Illustration: Google DeepMind

The first result is about digital signatures — how a computer checks that a piece of data, like a software update, really came from its claimed source. Before signing, most standards add preset data, called padding, to fill the message out into a fixed-sized block. Cryptographers knew padding helped prevent forgery, but not how much protection it offered. DeepMind's answer: what matters is how much of the signed block the attacker controls. Once they control roughly half of it or more, padding adds no protection at all — forging becomes as easy as if there were no padding.

The second result is about factoring many keys at once. RSA keys come in standard sizes, so an attacker who wants many keys can share one big up-front computation across all of them. DeepMind's agents found a new way to do that sharing — a "fan" of number fields fed by one shared sieve — that cuts the average cost of each additional 1024-bit key by about 30 times against the best batch method since 1993.

Two kinds of numbers

The announcement mixes two kinds of numbers, and keeping them apart is the whole story. One number was measured. The other two were calculated.

WhatFigureMeasured or estimated?
Reproducing the 1024-bit signature attack391 CPU core-years vs 1,380 — 3.5× less CPU timeMeasured, in a sandboxed simulation, on a key that was already broken
Forging a signature, 32-byte fingerprint~1,600× better than the previous best methodTheoretical estimate — the new algorithm was never run at 1024 bits
Factoring each extra 1024-bit key~30× cheaper — about 1/58 of a full factoring runLeading-order estimate from the cost formulas, not a run

The measured number is the reproduction. In September, an academic team — Laura Shea, Miro Haller, Adam Suhl, Heninger and Thomé — published a 1024-bit signature forgery that took 1,380 CPU core-years over about five months, asking a hardware security module to sign about 4 billion chosen values. DeepMind's two agents redid the same computation on the same key in 391 core-years. The paper is blunt: "the factor of 3.5 is one of implementation and execution, not of algorithms." The agents wrote about 33,500 lines of new code and debugged the run in under four days on roughly twenty short messages from a human researcher, most of them approvals. On up to 300,000 cores, the whole thing took 3.2 days from the first full-scale job, against about five months for the original.

The estimated numbers are arithmetic, not attacks. DeepMind's own essay calls the 1,600× a theoretical estimate, from the paper's cost formulas — the old method's constant 1.8171 against the new one's 1.5803. The 30× comes from the same kind of arithmetic on the batch-factoring formulas: the paper says the factors hidden in them "are not determined by the analysis" and that leading-order estimates "do not predict running times." Both are real mathematics. Neither is a demonstration.

Why it matters

First, what 1024-bit RSA is. RSA is the public-key system that has underpinned encryption on the internet for decades: multiplying two large primes is easy, reversing it — factoring their product — is hard. A 1024-bit key was the standard size in the 2000s; it is not now. NIST, the US standards body, stopped accepting 1024-bit RSA for new signatures in 2013, and almost everything in use today is 2048 bits or larger. This result is about a retired key size — and that is the point. Cryptographers use retired sizes as safe proving grounds, and modern encryption is unaffected.

Second, the speed. Major advances in this kind of work historically take years or decades, so defenders assumed the safety margins changed only slowly. This one took weeks — and the measured reproduction took days. The lesson is not that RSA is broken. It is that the cost of stress-testing a hardness assumption has collapsed — assumptions that looked safe for years can now be re-checked in weeks. That is the same clock our Quantum Computing page tracks: NIST's October 2026 schedule deprecates RSA-2048 and elliptic-curve algorithms by 2030 and removes them by 2035, because a future quantum computer running Shor's algorithm would break them outright. The DeepMind result is classical, not quantum, but it points the same way: the margins shrink, and the migration has a deadline.

Third, the oracle. The September paper's attack needs something specific: temporary access to a device that will sign chosen inputs without padding — a raw signing oracle. Standard signing interfaces, PKCS#1 v1.5 and PSS, do not generally expose one — the academic team had to switch off the FIPS-approved mode on its own hardware security module to get one. But the September paper's own conclusion is the part that should make the industry sit up: even 4096-bit RSA, it says, "does not appear to meet a 128-bit security level" in this attack model. That is a claim about the model, not about RSA as normally used — and it is why retiring old interfaces matters as much as retiring old key sizes. What actually gets systems broken is rarely the mathematics, which is the story our Cyber Security page keeps coming back to.

What it doesn't prove

It does not prove that RSA is broken. The papers say so themselves, in bold: nothing here threatens deployed cryptosystems based on 2048-bit or larger RSA. The biggest numbers are estimates — the batch paper is explicit that they do not predict running times. The measured 3.5× is an engineering saving on a reproduction, not a new algorithm, and the paper says so.

It does not prove that AI can do cryptanalysis on its own. The signature paper describes two cooperating agents — a supervisor that designed the attack and audited every result, and an executor that wrote the code and ran the jobs — with a human researcher sending about twenty short messages, most of them approvals. The batch paper's proofs were generated by one model and verified by a second, with the key identities machine-checked in Lean 4. Human experts chose the problems, set the directions and judged the results; the agents did the exploratory work. That is the pattern to watch: AI-assisted research, not autonomous discovery.

And it is not peer reviewed yet. Both papers are working papers, and no independent group had reported reproducing either result in the days after the announcement. The honest summary is not "AI broke RSA." It is: human-guided AI agents found two real weaknesses in a retired key size, measured one carefully, estimated the other two — and showed that checking the foundations of encryption just got a lot faster.

Sources & further reading

Checked November 13, 2026. Pushmeet Kohli, "Advancing the frontier of cryptanalysis," Google DeepMind, November 6, 2026 — linkedin.com. Dvijotham et al., "Modular Roots from a Restricted Oracle: Faster Affine Roots and a 1024-bit Computation," Google DeepMind working paper — deepmind-media (PDF). Dvijotham et al., "Fan Factories: Multiple Number Fields in Factorization Factories and Batch NFS," Google DeepMind working paper, November 5, 2026 — deepmind-media (PDF). Shea, Haller, Suhl, Heninger and Thomé, "Forging 1024-bit RSA signatures in nearly SNFS time," IACR ePrint 2026/2131, September 20, 2026 — eprint.iacr.org. The 1,600× and 30× figures are DeepMind's theoretical estimates, not measured results; the 3.5× is a measured reproduction. The ePrint record itself could not be fetched directly (access denied) and its details are corroborated from the DeepMind papers and independent coverage.

Google DeepMind and DeepMind are trademarks of Google LLC, shown here for identification only. No endorsement is implied. The hero image is an AI-generated abstract, not a photograph.

← Back to DTN News