Two results against a retired key size
Google DeepMind published two cryptanalysis results on Friday that it says were produced by AI agents working under human supervision, both against 1024-bit RSA — a key size retired from modern practice and kept by researchers as a safe proving ground. The lab said the systems “can be compromised with a lower computational budget than previously thought possible”.
The announcement came from Pushmeet Kohli, chief scientist at Google Cloud and a vice-president at DeepMind, who wrote that the results were led by Dj Dvijotham and Abhradeep Thakurta in collaboration with Nadia Heninger of UC San Diego and Inria and Emmanuel Thomé of Inria. Two technical papers were published alongside the post.
DeepMind’s own framing is about speed rather than any break. Advances of this kind have historically taken years or decades, the post says, which let defenders assume that the hardness margins underneath RSA moved slowly. The company’s claim is that an agent system compressed that work into weeks.
Padding stops helping once an attacker controls half the block
The first result concerns digital signatures — the mechanism that lets a computer check that a software update or a document really came from its claimed source. Before signing, many standards pad a message out to a fixed-size block. Cryptographers knew padding made forgery harder; what they did not have was a measure of how much protection it buys when an attacker can feed chosen inputs to a signing device.
According to DeepMind, what matters is how much of the signed block the attacker controls. Once that share reaches roughly half or more, the company says, padding adds no protection at all, and forging a signature becomes as exposed to mathematical shortcuts as signing with no padding.

The team tested that claim in a sandboxed simulation rather than against live systems. On an already-solved 1024-bit benchmark key taken from an earlier academic study, DeepMind says its agents engineered an implementation that reproduced the result using 3.5 times 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 — a short 32-byte digest, for example — forgery is harder, but the lab says it found a new algorithm that improves on the previous best method by roughly 1,600 times for 1024-bit keys. That figure is a theoretical estimate, not a measured run.
A cheaper second key, and a cheaper third
The second result is about factoring many numbers of the same size, which matters for RSA because keys are issued at standard bit lengths. An attacker going after many keys at once can in principle spread one large up-front computation across all of them. DeepMind says its agents found an algorithm with better asymptotic complexity than previously known batch methods, cutting the average cost of each additional 1024-bit key by about 30 times.
Both numbers are the lab’s own, and neither paper has been through peer review. The collaboration does, however, include two of the people who have run the field’s record factoring computations: Heninger and Thomé were part of the team that factored RSA-240 and RSA-250.

What it does not mean
Nothing here touches the encryption in ordinary use. DeepMind says the results affect only obsolete 1024-bit RSA, and that modern protocols, which almost universally use 2048-bit or larger keys, remain secure. The lab’s stated reason for publishing is that the security community should be able to use the insights to measure cryptographic strength and keep standards ahead of the threat.
The broader argument in the post is that computational hardness can no longer be treated as a fixed assumption that erodes on a predictable schedule. If an agent system can probe a hardness assumption in weeks, the interval between a theoretical advance and a practical one shortens for defenders as well as attackers.
What to watch is whether independent cryptanalysts reproduce the two algorithms, and in particular whether the 1,600-times estimate survives contact with an implementation. Both papers are public, so that work can start immediately.