From a Claude / Claude Code builder’s perspective, this is one of those stories that feels half thrilling, half mildly unsettling. Schneier is pointing at a new benchmark showing that frontier models are no longer just parroting cryptography lore; they’re starting to find real weaknesses, including some that researchers say were not previously known.
What strikes me is that this is not the usual fuzzy “AI understands math” claim. Cryptanalysis is a nasty domain with crisp outcomes: either the attack works or it doesn’t. That makes it a lot more believable as a benchmark than the usual vibes-based reasoning tests. If a model can help find a real flaw in an AEAD design or expose a proof mistake, that’s not benchmark theater. That’s material.
I also think the Claude angle matters here. The source says Anthropic’s model initially resisted going down the AES path, then researchers kept prompting, and after a lot of back-and-forth it refined the attack. That sounds very familiar to anyone who has worked with frontier models: the first response is often cautious, maybe too cautious, and the useful work only starts once you give the model a good harness and a persistent workflow. In other words, the model alone is not the story. The agentic setup is.
At the same time, I’d be careful not to overhype this into “LLMs can break modern crypto now.” The article is much more measured than that. The benchmark uses historical algorithms, scaled-down variants, and a frontier challenge set; the most dramatic results are still early signals, not a declaration that all contemporary cryptography is in trouble. The line about some attacks costing hundreds of millions of dollars is a useful reality check. Novel doesn’t automatically mean practical for attackers.
If I were building with Claude Code, I’d actually be interested in using something like this as a research assistant for protocol review, not as a magic cryptanalyst. I think the productive use case is: give the model a candidate design, ask it to search for weird edge cases, proof gaps, or reduction assumptions that feel brittle, then have humans verify everything. That feels much more concrete than expecting it to autonomously crack real-world systems.
The broader takeaway is pretty clear: Claude-class models are crossing into a part of security work that used to feel safely human-only. That doesn’t mean they replace cryptographers. It does mean the bar for “nobody would think of that attack” is getting higher, and faster than I expected.
Reference: Measuring LLMs' Ability to Perform Cryptanalysis - Schneier on Security
