AI can’t find an algorithm that doesn’t exist. The worry is we’re wrong about what exists. Liz Harkavy and Justin Thaler on AI vs. cryptography: "We generally believe there is no efficient algorithm for these problems. ... And so if there's not one, it doesn't matter how smart the... AI is; it cannot find the non-existent algorithm."show more

@liz_harkavy @SuccinctJT The bigger near-term risk is not AI solving factoring. It is AI finding bad parameters, weak implementations, and side channels faster than humans audit them. Assumptions fail at the edges first.

@liz_harkavy @SuccinctJT yeah the math doesn't care how smart the model is

@liz_harkavy @SuccinctJT maybe the real worry is what we haven’t

@liz_harkavy @SuccinctJT Exactly! We have not been getting any clue or trace on how AI can invent indpendently a more efficient algorithm to break ECC than Shor's. That's why we strongly disagree with Justin's remarks below @drakefjustin on ECC could be broken soon by AI.

@liz_harkavy @SuccinctJT so we're betting on P vs NP

@liz_harkavy @SuccinctJT right, if a model could break, say, RSA, that model would be an efficient algorithm for the problem. If no such algorithm exists, no set of weights can do it either

@liz_harkavy @SuccinctJT AI may accelerate implementation attacks before it changes the underlying hardness assumption. Protocols need separate controls for primitive selection, circuit correctness, key handling, and implementation leakage. One security label cannot cover all four.

@liz_harkavy @SuccinctJT the algorithm thing assumes p vs np stays unsolved, good luck with that

@liz_harkavy @SuccinctJT so you're betting on belief rather than actual validity in algorithms

@liz_harkavy @SuccinctJT This cuts both ways — if we're wrong about what exists, AI might be exactly what proves us wrong. Crypto's threat models assumed human-speed discovery.

@liz_harkavy @SuccinctJT The nearer-term risk is AI finding the implementation bugs humans already left in. The math holds; the code around it doesn't always

@liz_harkavy @SuccinctJT Yes but it can find relationships between algorithms, and take algorithms from one field and apply it to another. That's the difficult part and AI can do that well if steered properly