[Cryptography] "Anthropic A.I. Model Finds Flaws in Tough-to-Crack Encryption Algorithms"

[Jon Callas]

> On Jul 30, 2026, at 21:11, Jerry Leichter <leichter at lrw.com> wrote:
> 
> It's a faster attack on a weakened version of AES.  Also mentioned is some kind of an attack on a proposed post-quantum protocol (HAWK).

Reduced round, more than weakened. It's normal that a cipher gets analyzed in reduced round versions. 

This particular attack is not really a big deal, because it is reduced round, and also because it takes 2^105 data to do it. That's not going to happen in our space-time continuum.

> 
> Gift link to the NY Times article about this below.  Obviously the Times is not the place to learn any technical details.  Anyone come across a deeper description?

It's a real result, but it's not a big deal. We've known that AES 256 (and 192) are a bit weak. Even when Rijndael was selected as AES, this was a criticism. There's an existing attack that reduces the full thing to about 2^100, but also needing over 2^100 data -- I don't remember all the details, it's been quite some time. There was time when people opined that it meant that AES-256 was weaker than AES-128, which is kinda true and yet again not practical as no one is going to get that much data, nor really be able to do that many calculations even for one key.

I don't think that the attack is even as good as the existing one, given that it's a very reduced round attack and requires similarly absurd amounts of data. 

Nonetheless, it is a new, actual attack and thus a real result for AI-based cryptanalysis. Just not one that would be a publishable result if a human had done it. It's like the checkers-playing dog.

(The joke is that someone's walking through a park and sees a man playing checkers with a dog. He watches for a while and then exclaims, "That's amazing! I've never seen a dog playing checkers before!" The guy playing against the dog nods, and shrugs a bit and says, "Yeah, but he always loses." The AIs, like the dog, are going to get better over time. 

	Jon