Discoveries · 01
Where does AI stopcracking lattice crypto?
AI can break simplified post-quantum encryption when the puzzle is easy. Nobody has drawn the line where easy ends. Drawing it is the research.
How this was found
Our system read about 4,800 papers, citing every claim, and kept pointing at one thing: attacks that work inside a small box, and no map of where the box ends.
Feel the problem
Kyber and Dilithium rest on a problem called LWE. It starts as school algebra. Two secret numbers hide behind these equations. Try to find them.
The real secret (the solver can’t see it): x = 2, y = 5
That is LWE: equations plus a little noise. Real systems use hundreds of unknowns. No known method solves that, unless the puzzle is easy enough.
The unmapped boundary
Two dials: how many unknowns (n) and how many are non-zero (h). Known attacks win bottom-left. Real encryption lives top-right. Move the dials.
Unmapped. No published result either way. This is the ground we are charting.
A sketch, not data. Published attacks stop near n ≤ 128 with very sparse secrets.
The open question
Where is the line between “AI cracks it” and “safe”? Does it move with more data and more GPUs? Drawing it, with honest error bars, shows how much margin the standards really have.
To be clear
A lab crack is not a real-world break
A toy result with one method and one budget says nothing about real encryption.
Published attacks are far from real settings
The best results stop near n ≤ 128 with sparse secrets. Kyber and Dilithium use n = 256 to 1024 with dense ones.
Nothing here weakens Kyber or Dilithium
This measures the margin. It does not shrink it.
Built on
The papers this starts from.
…and later work pushing toward real-world settings.