Skip to main content

Can AI Break RSA Encryption?

We’re told that it will take powerful quantum computers to break RSA encryption, so for now the world is safe. But I wondered, in an era of increasingly sophisticated models, might AI pose a threat? These systems excel at finding patterns in data that humans miss, and if there were any subtle weaknesses in key generation, I would think AI could detect them.

Yes, theory says it’s all but impossible to break RSA because it relies on the computational hardness of factoring large prime numbers. But theory and practice don’t always align, and sometimes the most interesting discoveries come from testing our assumptions. So I set up an experiment to test whether a transformer model could learn to reverse-engineer SSH private keys from their corresponding public keys.

Experiment and Results

I trained a T5-small transformer model (60 million parameters) on a dataset of 50,000 SSH key pairs, split into 70% training, 15% validation, and 15% test. Given a public key as input, the model was asked to output the corresponding private key.

For hardware reasons, I decided to work with a smaller model (I could provision more resources and try a larger model if the results indicated something interesting). As it was, the experiment ran “comfortably” on my Mac M1, with 1 epoch taking about 15 hours.

In the first 25% of training, the model quickly learned. Loss dropped from 7.66 to 4.58. I suspected it was learning superficial structure, such as base64 encoding patterns and the standard SSH private key file header and footer. But the loss soon plateaued, stabalizing around 4.56-4.54, with only tiny improvements during the remainder of training. The validation loss showed similar behavior, decreasing from 4.52 to 4.49 over the final 50% of training. this suggests the model had hit a barrier.

For RSA-2048, the probability of randomly guessing a private key is approximately 2-2048, which is essentially zero. So, this sort of stagnation is exactly what we’d hope to see. The model had reached the the point where it had exhausted all learnable patterns except the actual mathematical relationship.

When I tested the trained model on unseen public keys, the results were reassuring. The model generated outputs that were structurally correct but cryptographically invalid. The first ~80 characters were identical, but this is the OpenSSH private key header. Where the actual cryptographic material begins, the outputs diverged.

Conclusion

I’d like to think this experiment provides empirical validation of RSA’s security against pattern-based attacks. A 60-million parameter transformer, trained on 50,000 unique examples, could not find any exploitable patterns in SSH key generation. And the plateau suggests further attempts with a larger model or more training would continue to be unsuccessful. The model did indeed learn formatting, but failed at the cryptographic content.

So much for any interesting discoveries. At least the world is indeed safe … at least for now.

Code for this experiment is available on GitHub.

Comments

Popular posts from this blog

What's Wrong with OOP

Proponents of Object Oriented Programming feel the paradigm yields code that is better organized, easier to understand and maintain, and reusable. They view procedural programming code as unwieldy spaghetti and embrace OO-centric design patterns as the "right way" to do things. They argue objects are easier to grasp because they model how we view the world. If the popularity of languages like Java and C# is any indication, they may be right. But after almost 20 years of OOP in the mainstream, there's still a large portion of programmers who resist it. If objects truly model the way people think of things in the real world, then why do people have a hard time understanding and working in OOP? I suspect the problem might be the focus on objects instead of actions. If I may quote from Steve Yegge's Execution in the Kingdom of Nouns : Verbs in Javaland are responsible for all the work, but as they are held in contempt by all, no Verb is ever permitted to wander about ...

Does Esperanto-USA have a Transparency Problem?

I’ve been an active member of Esperanto-USA for many years, even serving on its board of directors. I wanted to help build a strong, healthy national organization that effectively promotes Esperanto in the United States. But when I stepped down from the board, that desire didn’t end. I continued to bring problems to the attention of the organization, ranging from member rights to strategy to proper governance. I’ve decided to be public about this latest issue because of its seriousness and because of how the board has been secretive about it. The membership deserves to know what is done in their name. What Happened In November 2025, the board of Esperanto-USA sent a formal letter to Esperanto-Ligo en Israelo (ELI) with a copy to UEA, concerning an article published in ELI’s bulletin. The article, written by Eduardo Topol, contained inflammatory political content. And although I agree that the article was extremely offensive, the board’s decision to inj...

Learning Prolog

I'm not quite sure exactly I was searching for, but somehow I serendipitously stumbled upon the site learnprolognow.org a few months ago. It's the home for an introductory Prolog programming course. Logic programming offers an interesting way to think about your problems; I've been doing so much procedural and object-oriented programming in the past decade that it really took effort to think at a higher level! I found the most interesting features to be definite clause grammars (DCG), and unification. Difference lists are very powerful and Prolog's DCG syntax makes it easy to work with them. Specifying a grammar such as: s(s(NP,VP)) --> np(NP,X,Y,subject), vp(VP,X,Y). np(np(DET,NBAR,PP),X,Y,_) --> det(DET,X), nbar(NBAR,X,Y), pp(PP). np(np(DET,NBAR),X,Y,_) --> det(DET,X), nbar(NBAR,X,Y). np(np(PRO),X,Y,Z) --> pro(PRO,X,Y,Z). vp(vp(V),X,Y) --> v(V,X,Y). vp(vp(V,NP),X,Y) --> v(V,X,Y), np(NP,_,_,object). nbar(nbar(JP),X,3) --> jp(JP,X). pp(pp(PREP,N...