AI Won’t Crack Your Encryption. It’ll Build Something Worse.

You’ve probably seen the headlines. Some AI researcher claims GPT-7 will crack RSA by Christmas. Security Twitter panics. Then nothing happens. Again.

Here’s the thing: the people saying LLMs will break your encryption are wrong. Dead wrong. And the people saying LLMs pose zero threat to cryptography are also wrong — just in a more boring, more dangerous way.

A next-word prediction engine can’t factor two insanely large primes. That’s not a limitation. It’s a distraction from where the real damage is being done.

Let’s get the obvious out of the way. Symmetric cryptography — AES-256, the stuff protecting your bank transactions and your Signal messages — is mathematically brutal. There’s no pattern to exploit, no shortcut to discover, no clever prompt that makes the model suddenly intuit the answer to a problem that requires more brute-force attempts than there are atoms in the universe. An LLM predicting the next token in a sentence has about as much chance breaking AES-256 as a goldfish has of solving a Rubik’s Cube.

One commenter on the original analysis nailed the obvious: “A next-word-in-the-sentence prediction engine can’t predict the factor of two insanely large prime numbers… tell me more, captain obvious!”

And yeah, fair. If the story ended there, this would be a boring article about how AI can’t do something it was never going to do. But the story doesn’t end there.

Here’s where it gets uncomfortable.

While everyone’s debating whether ChatGPT can crack AES directly, LLMs are quietly doing something far more insidious: they’re accelerating the research that will eventually make today’s encryption obsolete. Not by attacking the lock. By building a better key.

Think about what’s actually happening in quantum computing labs right now. Researchers are using LLMs to parse decades of physics literature in hours. They’re using them to optimize quantum error-correction algorithms, to identify patterns in quantum noise data that humans would take years to spot, to generate and test hypotheses at a pace that makes traditional peer review look like snail mail.

The threat was never that AI would crack your code. The threat is that AI makes the person who cracks your code ten years smarter, ten years faster.

Shor’s algorithm — the quantum approach that theoretically breaks RSA and ECC — has existed since 1994. What’s prevented it from being practically deployed isn’t the math. It’s the engineering. We need quantum computers with thousands of stable logical qubits, and we’ve been stuck in the dozens. But the gap between “dozens” and “thousands” is a research problem, not a theoretical one. And LLMs are extraordinarily good at accelerating research problems.

Another commenter on the original piece saw this clearly: “LLMs will accelerate math research, increasing understanding in areas like quantum which will eventually lead to breakthroughs that will break most standard asymmetric encryption algorithms.”

That’s not speculation. That’s trajectory.

But there’s a third vector that nobody’s talking about. The most underrated comment from the source analysis pointed out that cryptographic systems rest on three pillars: mathematical impossibility, brute-force time requirements, and correctness of implementation. LLMs can’t touch the first two. But the third? The code and algorithms themselves?

That’s where LLMs are already in the game. Not by breaking crypto, but by writing the implementations that run it — and introducing subtle, nearly undetectable vulnerabilities along the way. A misplaced constant. An off-by-one in a key schedule. A timing side-channel that leaks your private key to anyone patient enough to measure response times.

One commenter put it precisely: “Cryptographic systems are based on 1) mathematical impossibility of reversing some integer/mod calculation, 2) time required for a brute force attack, 3) correctness of algorithms and code used in implementations. The last part is where LLMs have a chance.”

So let’s reframe the threat model. Your data isn’t safe because AI can’t break AES. Your data is safe today because the specific path from “LLM exists” to “your encryption is broken” hasn’t been walked yet. But the path exists, and it runs through quantum computing research, through algorithmic optimization, through implementation bugs in the very code that LLMs are increasingly writing for us.

They won’t pick the lock. They’ll build the key that makes the lock irrelevant — and they’ll do it while you’re busy arguing about whether they can pick locks at all.

If you work in security, this changes your timeline. Post-quantum cryptography migration isn’t a 2030s problem anymore. It’s a problem whose deadline is now being accelerated by tools that can compress years of literature review into an afternoon. Every quantum computing breakthrough that LLMs help enable is a tick closer to the day when RSA-2048 goes from “military-grade” to “museum exhibit.”

And if you’re just someone who relies on encrypted communications — which is everyone with a phone — you should feel two things right now. Relief, because today’s encryption holds. And unease, because the thing that will eventually break it is already running, quietly, in a research lab somewhere, accelerating the exact science that makes your secrets expirable.

The question was never “Can AI break encryption?” The question was always “How fast can AI help someone else do it?”

And that answer, unlike your AES-256 keys, is changing every single day.