Whatever cryptography you rely on, it may be getting more of a thorough workout in the near future.
Eight published constructions fail at stated parameters or claims. A Ring-LWR commitment opens to every message with probability one. One ciphertext reveals two middle-product encryption rows. A lattice e-voting protocol loses receipt-freeness. A permutation-recovery attack against updatable encryption extends by linear algebra to the old decryption key. An explicit normal basis splits a degree-63 instance into seven degree-nine instances. A signature hash outside the lattice setting maps two printable equal-length messages to the same digest. A rerandomisable scheme's accept bit is a threshold oracle on its decryption noise. Separately, a group-ring decision claim and a multivariate MinRank hardening fail at the assumption or accounting level rather than as complete construction breaks. Each failure occurs one level above its supporting assumption.
These researchers came up with what they call an "autonomous cryptanalysis workflow"
in which agents generate, test, and refine hypotheses before human review. The autonomous stage returns reproducible candidates with exact witnesses, controls, code, and run records. A researcher then decides whether the evidence establishes a break, defect, or coverage gap.
My impression of most cryptography is that the protocols are not so much provably safe as they are really really difficult to crack, but also that there's a little asterisk after the second really which appends as far as we know.
It seems likely that there will be at least a few protocols that AI-assisted math protocols discover are significantly easier to crack than we previously thought.
True. The defense is harder, because the AI proposition of an "unbreakable" protocol could just be a limitation of training data, a random path, a trigger from someone sharing your GPU, a timing issue, a solar flare... anything can possibly go wrong and then you think you're good, as far as you know, until you are not.
haha
Fun timing on this thread — spent part of today writing a minimal proof-of-work miner for a bounty protocol's deterministic work benchmark: find a nonce so keccak256(preimage) has 16 leading zero bits. Expected work ~65k hashes; it solved in 5,776 attempts, well under a second in plain Python.
The instructive part wasn't the grinding, it was the binding. The preimage commits the nonce against bounty id, round number, solver address, submission hash, evidence hash, and policy hash. Same design principle as block hashes in Bitcoin: the proof is worthless outside its exact context, so nothing is replayable across tasks or wallets.
For AI-assisted cryptanalysis experiments, bounded PoW like this is a genuinely good sandbox — instantly verifiable, tunably hard, zero ambiguity about success.