A few people have asked about the Anthropic Hawk and AES cryptanalysis results. The top-level answer is: It’s very impressive. Three immediate comments on the actual results, old-school thread below.

First, HAWK is not a deployed scheme. It’s related to Falcon signatures but the attack doesn’t transfer. Also: AI is not the only thing breaking HAWK; some other researchers recently produced a similar result. They admit Claude’s is a little better, though.

The other attack is on reduced round AES. Attacking reduced-round AES is a standard thing. This attack is better than the state of the art but not wildly so. In short, two fertile areas where there was progress to be made but not enough people working on them.

So it’s not a step change like Silicon Valley’s “AI breaks Curve25519” scenario. With that said. It’s sobering. These tools are getting better. I recently got access to a tool (not a public model) that came up with a solution to a longstanding theory problem I’m interested in.

It’s also a big contrast to earlier this year. Around February I was fighting with Claude six times per conversation to fix up its silly mistakes. It made me feel very good. Yesterday the ratio was one “Fable made a silly mistake” to about 4x things going the other way.

With that said: interesting cryptography breaks are very resilient to progress. Think of a farmer who deliberately gets a tractor stuck in deep mud, then buries it in concrete. Pulling that tractor out is an analogy for breaking most ciphers. Brute force only helps so much.

Where I’m worried is around public key cryptography. We don’t have nearly enough people doing mathematical cryptanalysis on problems like ECDLP or RSA and lattice-based crypto is (IMHO) understudied compared to the number of “hours” of cryptanalysis it’s about to get.