DeepMind AI Agents Find 2 Weaknesses in Legacy 1024-Bit RSA

Google DeepMind says its AI agents, “guided by human supervision”, found two weaknesses in legacy 1024-bit RSA, the public-key system long used to sign and encrypt data. Pushmeet Kohli of Google DeepMind set out the results in an essay on X on 6 November: “these systems can be compromised with a lower computational budget than previously thought possible.”

The essay is explicit about the limit: “In practice, these results only affect obsolete 1024-bit RSA systems, and modern encryption protocols, which almost universally use 2048-bit or larger keys, remain secure.” Of the two headline gains, the 1,600x figure for signature forgery is, in the essay’s words, a “theoretical” estimate; the one result the essay describes as run in practice used 3.5x less CPU time on a key that had already been solved.

Business Pill · AI Agents Explained

A short explainer of how AI agents work. It teaches the general idea only and does not discuss Google DeepMind, RSA or this research.

Result One: Signature Padding

Digital signatures prove that a message, such as “a software update, document or a website”, comes from its claimed source. Before signing, many standards add “padding” to fill the message out to a fixed-size block. The essay says the team asked how much protection padding gives when an attacker can feed chosen inputs to a signing device.

Its answer: “what matters is how much of the signed block an attacker controls. Once they control roughly half of the block or more, padding adds no protection at all.” In a “sandboxed simulation—without touching any live systems”, the essay says, the agents worked on “an already-solved 1024-bit benchmark key from an earlier academic study, reproducing the result with 3.5× less CPU time, and at less than a thousandth of the computing power needed to break the key itself.”

Where an attacker controls less of the block, “like with a short 32-byte digital fingerprint in the message”, forgery is harder. Even so, the essay says, the team found “a new algorithm that, according to our theoretical estimates, improves on the previous best method by roughly 1,600x for 1024-bit keys.”

What Was Posted on X

Pushmeet Kohli of Google DeepMind announced the two results on X on 6 November, saying they do not affect modern schemes such as RSA 2048.

An X user summarised the results as a roughly 1,600x gain for signature forgery and a roughly 30x cut in the cost of factoring additional keys in large batches; the essay calls the 1,600x figure a theoretical estimate.

Result Two: Factoring Keys in Bulk

RSA keys usually share a standard length, such as 1024 or 2048 bits. The essay explains that an attacker breaking many keys “can spread the cost of one large up-front computation across all of them, lowering the cost per key.” It says the agents found “a new algorithm with better asymptotic complexity than previously known methods, cutting the average cost of factoring each additional 1024-bit key, out of many, by about 30× compared to previous batch methods.”

Who Did the Work

The essay says the results were “led by Dj Dvijotham and Abhradeep Thakurta and in collaboration with Nadia Heninger (UC San Diego and Inria), Emmanuel Thomé (Inria)”. In a post on X, Kohli wrote that the team “showed how AI agents could make two significant cryptographic advances in just weeks”. The essay says such advances have “historically” taken “years or decades”.

The essay refers readers to “our full technical papers on digital signatures and key factorisation”. We searched arXiv and the IACR ePrint archive on 6 November and did not find them, so we have not read the papers or their proofs.

Why DeepMind Says It Matters

The essay’s case is about assumptions, not live systems: “in a new era of AI-enabled mathematics and computation, we cannot treat cryptographic hardness as a fixed assumption.” Kohli’s post adds that the results “show why we must not assume the mathematical “hardness” of cryptographic schemes which we have relied upon in the past.”

Connecting the Coverage

Earlier: PSI Says Anthropic Model Proved 18-Year Quantum Conjecture, OpenAI’s GPT-6 Cyber Hits 48.5% on ExploitGym, Up From 32.1%, OpenAI’s AI Reproduced 1,724 Linux Kernel Flaws in $1.2M Run and Claude Mythos Preview Finds a Real Weakness in HAWK. What is new here: a frontier lab says its agents produced two cryptanalysis results on a widely studied public-key system, while stating that modern 2048-bit keys are unaffected.

The Structural Read

As we read the essay, the signature result is about who controls the signed block: it says that once an attacker controls roughly half of it, padding adds no protection.

As we read the essay, the only result it describes as tested ran in a sandboxed simulation on an already-solved benchmark key, with 3.5 times less CPU time.

As we read the essay, the team includes outside cryptographers, Nadia Heninger and Emmanuel Thomé, alongside DeepMind’s Dj Dvijotham and Abhradeep Thakurta.

Three Implications

FOR SECURITY TEAMS The essay says the results only affect obsolete 1024-bit RSA systems and that modern protocols using 2048-bit or larger keys remain secure.

FOR AI RESEARCH CLAIMS Kohli wrote that the agents made the advances in weeks; the essay says such advances historically took years or decades.

FOR READERS The technical papers were not located on arXiv or the IACR ePrint archive when we searched; we found no independent assessment.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework says: “The surface, not the model, governs autonomy.”

As we read it, the essay puts weight on the setup around the model: it says the agents worked under human supervision and in a sandboxed simulation. The framework does not discuss Google DeepMind, RSA or cryptography.

What Is Not Established

The results come from Google DeepMind’s own essay and posts. We did not read the technical papers, and we found no assessment from independent cryptographers. The 1,600x forgery figure is a theoretical estimate, and the 30x factoring figure applies to each additional key in a batch, not to breaking a single key. We did not contact Google DeepMind.

The Bottom Line

Google DeepMind says AI agents found two weaknesses in obsolete 1024-bit RSA: a padding result for signatures, with a theoretical 1,600x gain in one setting and a 3.5x CPU-time saving reproduced on an already-solved key, and a batch-factoring method about 30x cheaper per additional key. It says 2048-bit and larger keys remain secure.