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

What DeepMind claims

Google DeepMind says AI agents working under human supervision have turned up two weaknesses in legacy RSA cryptography built on 1024-bit keys. According to a dev.to write-up of the announcement, the findings were disclosed on November 6, 2026, and documented in two working papers published on November 5 and 6: "Modular Roots from a Restricted Oracle" and "Fan Factories." Named co-authors include Dj Dvijotham, Abhradeep Thakurta, Nadia Heninger, Emmanuel Thomé and Pushmeet Kohli, among others. One caveat matters up front: neither paper has been peer reviewed, and the figures below are the researchers' own claims as relayed by dev.to.

A cheaper route to forging signatures

The first weakness sits in digital signatures. When an attacker already controls a small slice of the data being signed — say a 32-byte fingerprint — the researchers estimate the new approach outperforms the previous best method against 1024-bit keys by a factor of roughly 1600. dev.to stresses two limits on that number: it is a theoretical estimate rather than a demonstrated attack, and the entire investigation ran inside an isolated environment.

Cutting the cost of factoring in bulk

The second weakness concerns factoring, the operation RSA's security rests on. Here, each integer to be cracked is assigned its own family of number fields. Once the budget for a first factorization has been spent, dev.to reports, breaking the next 1024-bit integer becomes about 30 times cheaper than it would be with the multiple number field sieve (MNFS), or roughly 1/58 of what a single run of the general number field sieve costs.

A reproduction, with machines checking machines

Beyond the two new results, the agents reproduced a 1024-bit computation that had been carried out before. The first root consumed 391 CPU core-years, about 3.5 times less than the original effort, and the full-scale sieving phase finished in 3.2 days using up to 300,000 CPU cores and 16 GPUs, where the earlier computation had needed around five months. The workflow itself is notable: one agent produced the mathematical proofs, and a second agent verified them in Lean 4. Along the way, the agents surfaced and fixed four problems affecting the algorithm or the correctness of the output, none of which required another query to the oracle.

What is and isn't at risk

DeepMind states plainly that current systems using keys of 2048 bits or larger are not affected by these results, and that nothing in the work touched live systems. 1024-bit RSA has been treated as legacy for years, and the findings sharpen that assessment rather than overturn it. Still, given the single secondary source and the absence of peer review, independent cryptanalytic review will be the real test of the claims.

Why it matters

Two things lift this beyond an academic exercise. First, the structure of the work: agents proposing mathematics while other agents machine-check the proofs compresses a research cycle in computational number theory, a field where progress normally takes years. Second, the planning implication: if the cost of breaking a given key size can drop faster than assumed, then key strength is not a fixed property but a moving target, and deprecation schedules for older algorithms deserve shorter review intervals. For engineers, today's practical advice is unchanged — no supported system should be running 1024-bit RSA — but how quickly that advice needs revisiting just became harder to predict.