Did AI agents just break RSA?

Short answer: no. Long answer: one headline, two very different results, and a real surprise in the second.

The headline you saw: Google DeepMind says its AI agents found two weaknesses in RSA. Cue the panic. Here is the calmer version. The weaknesses are in old 1024-bit keys, the kind security folks have been telling everyone to retire for over a decade. Modern 2048-bit keys and larger are not affected, and DeepMind says so itself.

Fine, but "no panic" is not the same as "nothing to see". There is something to see, it is just not where the headline points.

Result one: forging signatures

This one sounds scary: a cheaper way to fake RSA signatures. But look closer. A team of academics published a practical attack of this kind back in September, and it took them 1,380 CPU core-years. The DeepMind agents reproduced that same computation in 391. Same algorithm, better engineering.

That is impressive as a speed-up, and a decent demo of what agents can do when they are given a hard, well-defined optimisation job. It is not a new attack. It also only works when the attacker can get a signing oracle to play with, so it bites systems that sign blindly, not your average website.

Result two: factoring keys in bulk

This is the genuinely new bit. The agents came up with a method that makes factoring many 1024-bit keys at once much cheaper. Once you have paid for the first key, each extra one comes out about thirty times cheaper than with the best previous batch approach.

And now the obvious question: if these keys were retired years ago, who cares?

More people than you would think. 1024-bit keys are officially out for creating new signatures, but they are still allowed for checking old ones, and they still hide in legacy hardware, old certificates and forgotten corners of email and code signing. Cheaper batch attacks turn "nobody would bother" into "somebody might".

So what is the real story?

The attacker is changing. For decades, breaking a cipher meant a clever human plus a very large computer. Now it can be a supervised agent that tries ideas at machine speed, and DeepMind says this took just weeks, working with cryptanalysis experts. Not an autonomous genius. A human-guided search that moves faster than the field is used to.

The takeaway is boring in the best possible way. Find your old 1024-bit keys. Replace them. And keep your eyes on the next round of results, because the people stress-testing encryption just got a very fast new colleague.