Claude Mythos has achieved a major milestone for AI-assisted cryptanalysis, discovering new attacks against the HAWK post-quantum signature scheme and a reduced-round version of AES-128. The findings were independently verified by cryptographers, making this more than just an AI-generated hypothesis. This does not break real-world AES. Full AES-128 remains secure, and the attack only targets the 7-round research variant. The significance lies in AI independently producing novel cryptanalytic techniques that improve on previous work. As AI models become more capable, they could play an increasingly important role in evaluating the security of both classical and post-quantum cryptography.
I would be pissed at these headlines, if this were my work.
I, an expert, used the latest bullshit brute force libraries in a novel way to move human knowledge forward.
Headline: “Latest bullshit brute force library makes discovery!”
Soap box time:
Mythos is a clever algorithm, but it won’t make it into the history books.
The people getting excited have no fucking clue what is going on.
Claude Mythos is a VW Beetle - genuinely interesting design with an advertising budget larger than some counties, decorating a solution that isn’t remotely a good use of the money being spent on it.
My favorite comparison is chimpanzees…
A chimp’s working memory is absolutely insane compared to a human.
Flash a 50 symbol string to a chimp for a fraction of a second and give him a banana, and that chimp will walk back up to the screen and type in the same sequence for another banana as soon as they eat the first.
No human on the planet can do that, if that’s the comparison we’re all idiots and chimpanzees are the dominant species…
But there’s more to it than that. It’s just a very narrow slice where a chimp is “smarter” than a human.
AI can examine code and find errors/vulnerabilities/whatever, way faster than a human.
But that doesn’t mean it can do anything else better than a human.