From a Claude / Claude Code developer’s point of view, this is one of the more serious “wait, really?” demos Anthropic has published in a while. It’s not about writing code faster or summarizing papers; it’s about a frontier model helping uncover weaknesses in cryptographic algorithms themselves, including one post-quantum signature scheme and a reduced-round version of AES.
What strikes me is that this is exactly the kind of capability I’d want models to have in security work, and exactly the kind that should make us careful. Cryptography is built on public scrutiny. If an LLM can meaningfully accelerate that scrutiny, that’s a real contribution, not just a flashy benchmark win.
I think the HAWK result is the headline here, even though Anthropic is careful not to oversell it. It doesn’t mean “post-quantum crypto is broken” and it doesn’t mean Claude can go around cracking deployed systems. But it does mean a model can find a mathematically meaningful weakness in a candidate scheme that had already survived expert review. That’s impressive, and a little uncomfortable in the best possible way.
The AES result feels more familiar to anyone who’s followed academic cryptanalysis. Reduced-round attacks are normal research territory, and the source is explicit that this does not touch full AES. Still, seeing a model improve that line of work is interesting because it suggests these systems may be good at grinding through the combinatorics and literature-heavy parts that humans find tedious. That’s the part I’d actually want to test myself: not “can Claude crack AES,” but “can it help me explore a giant attack surface faster than I can alone?”
If I were building with Claude or Claude Code, I’d take this as a signal to try it on more bounded, high-signal research tasks: literature review, theorem chasing, generating candidate attacks, verifying edge cases, and building exhaustive test harnesses. I’d be more skeptical of grand claims that a model can “replace” cryptographers. This story reads more like augmentation than replacement. That’s exciting enough on its own.
I’d also be curious whether this kind of work generalizes beyond cryptography into other mathematics-heavy domains where the search space is huge but the rules are crisp. Perhaps that’s where models like Mythos become most useful: not as magical inventors, but as tireless adversarial assistants that notice the one symmetry humans missed.
The broad takeaway is simple: Claude is no longer just writing code around security systems. In at least some settings, it can help break the math those systems are built on. That deserves attention, and probably a little humility.
