Discoveries · 01
Where does AI stopcracking lattice crypto?
Machine learning already breaks simplified versions of post-quantum encryption — in easy cases. Nobody has mapped where “easy” ends and “safe” begins. Drawing that line is the research.
How this was found
Our autonomous discovery platform — grounded literature search over roughly 4,800 papers, answering only with traceable citations — kept pointing at the same corner: a line of transformer attacks on lattice cryptography that succeeds inside a small box, and a boundary around that box that no one has charted.
Feel the problem
Kyber and Dilithium — the new post-quantum standards — rest on a problem called LWE: Learning With Errors. It starts as schoolbook equations. A hidden secret — two numbers — sits behind the equations below. You get the equations, not the secret. Try to recover it.
True secret, hidden from the solver: x = 2, y = 5
That is LWE: equations plus a whisper of noise. Real deployments stretch it to hundreds of unknowns in clock arithmetic, where numbers wrap and size tells you nothing. No known method — classical or quantum — crosses that gap. Unless the puzzle is easy enough. Which is the point.
The unmapped boundary
Two dials set the difficulty: dimension n — how many unknowns — and secret weight h — how many of them are non-zero. Published transformer attacks win only in the bottom-left corner. Deployed encryption lives far top-right. Move the dials.
Unmapped. No published result says whether AI cracks this — exactly the territory this research charts.
Illustrative, from published results — the hardest reported recoveries stop near n ≤ 128 with very sparse secrets. The exact edges are precisely what nobody knows.
The open question
Across dimension, secret weight, and compute budget: where exactly is the line between “AI cracks it” and “safe” — and is it a fundamental wall, or does it move outward with more data and GPUs? Mapping that line with honest error bars and false-alarm-controlled claims tells you how much safety margin the new standards really hold against AI-powered attacks. That map is what we are building.
Staying honest
A lab crack is not a real-world break
Success at a toy setting is a lower bound under one method and one budget. It says nothing about deployed encryption.
Published attacks sit far below deployed parameters
The strongest reported recoveries stop around n ≤ 128 with very sparse secrets. Kyber and Dilithium run n = 256–1024 with dense secrets.
Nothing here weakens Kyber or Dilithium
This work measures the margin. It does not shrink it.
The line of work
The established results this builds on.
…plus follow-on work pushing these attacks toward the exact settings the standards use.