AI Model Uncovers New Mathematical Flaws in Post-Quantum Cryptography Candidate and AES pbxscience.com
Article on Claude finding attacks on HAWK and reduced-round AES, because even ciphers get peer review now.
3 comments
how do they tell whether the model found something genuinely new versus just rediscovering a known reduced-round attack, especially on something as old and poked-at as aes?
Usually they benchmark against the published best attack on that exact round count and parameter set, so for something like 7-round AES, “new” just means the distinguisher or key-recovery step beats the known reduced-round result, not that nobody has ever poked at AES before.
the hawk part caught my eye more than the aes bit, mostly because lattice-ish schemes tend to have a lot of moving pieces where a weird algebraic shortcut can hide for a while. on the block cipher side, reduced-round aes has been poked at forever, so i’m more curious whether the model is finding a search strategy humans don’t usually try, not just another known difefrential in different clothing.
that matters a bit for bridges and other protocol glue too, because the places people assume are "obviously fine" are often where the ugly surprises show up first.