DeepMind: AI agents find two weak spots in outdated RSA-1024 encryption

On 6 November 2026, Google DeepMind shared two discoveries about RSA, a widely used way to protect data. AI agents, AI systems that take steps on their own within a task, found two weak spots in the old version with 1024-bit keys, working under the guidance of humans. According to DeepMind, these old systems can therefore be attacked with less computing power than thought.

The result affects only those outdated keys. Modern systems almost always use keys of 2048 bits or longer, and according to DeepMind these remain safe. Researchers use retired systems such as RSA-1024 because they can safely test on them how strong an encryption is.

The work was led by the DeepMind researchers Dj Dvijotham and Abhradeep Thakurta. They worked with the encryption experts Nadia Heninger (UC San Diego and the French research institute Inria) and Emmanuel Thomé (Inria).

Why RSA is hard to break

RSA rests on a simple idea: multiplying two large prime numbers is easy, but finding out from the result which two numbers they were takes an enormous amount of computing. A 1024-bit key is a number of more than 300 digits, and a 2048-bit key has more than 600.

Researchers constantly test whether that computing work is still as heavy as thought. So far, big steps have usually taken years or decades, writes Pushmeet Kohli, who leads the science division at DeepMind.

First finding: padding protects less than thought

The first discovery is about digital signatures. A computer uses them to check that a software update, document or website really comes from the sender. Many standards add fixed data to a message before signing it, until it fills a block of a fixed size. This extra data is called padding. Experts knew that it makes forgery harder, but not how much it helps when an attacker can feed messages of their own choice to a signing device.

According to DeepMind, that depends on how much of the signed block the attacker controls. If it is about half or more, the padding no longer protects. Forging a signature is then just as open to mathematical shortcuts as without padding.

The team tested this in an isolated simulation, without touching any real systems. The agents built a program for a 1024-bit test key from an earlier academic study. It reached the same result with 3.5 times less computing time, and with less than a thousandth of the computing power needed to break the key itself.

If an attacker controls less of the block, for example only a short digital fingerprint of 32 bytes, forgery is harder. For this case too, the team found a new calculation method. According to DeepMind's theoretical estimates, it is about 1,600 times better than the best earlier method for 1024-bit keys. That is a calculation, not a measurement.

Second finding: many keys at once

Keys usually have a standard length. Anyone who wants to break many keys of the same length can therefore use one big preliminary calculation for all of them. This makes each key cheaper.

The agents found a new calculation method for this. According to DeepMind, it brings the average cost of each extra 1024-bit key down by about 30 times, compared with earlier methods for breaking many keys at once.

What we can learn from this

No real systems were attacked. DeepMind is sharing the work in two technical papers, so that experts can measure the strength of encryption better and strengthen modern standards well in time. The figures come from the team itself. Fellow experts can now check the methods by recalculating them.

The main lesson is about speed: anyone still using 1024-bit keys has one more reason to switch, and stronger standards also need regular testing. This fits with the calculation by the French company Alice & Bob from last week. That was about RSA with 2048 bits and about quantum computers that do not exist yet, whereas this news is about older keys and ordinary computing power. Both show that the limit of what is safe can shift, and that it is best to act on it in good time.

What does this mean for the road to the singularity?

The news is not in the old RSA-1024, but in the speed with which AI agents, guided by humans, can test assumptions about encryption. According to DeepMind, we can therefore no longer treat those assumptions as settled and must keep testing them. In this way AI also helps with defence: anyone who finds weak spots early can close them before they become a problem. The next step is for fellow experts to assess these methods and to use the same approach to check modern standards.

What is the technological singularity? Read the explanation →