Quoting Matthew Green
Simon Willison's Weblog Simon Willison
Cryptographer Matthew Green weighed in on Anthropic's new AI cryptanalysis research. Timing matters: we're mid-switch to post-quantum crypto, so AI muscle here could be a huge win or risk.
Based on reporting by Simon Willison's Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Matthew Green, the Johns Hopkins cryptographer known for keeping a close eye on where AI and security collide, has some thoughts on Anthropic's latest cryptography research, and they're worth sitting with. His framing isn't about doom or hype. It's about timing.
We're currently in the middle of a once-a-generation shift in public-key cryptography, moving away from the RSA and elliptic-curve systems that have underpinned the internet for decades and toward post-quantum alternatives built on entirely different mathematical problems. That's why proposals like HAWK are getting so much attention from standards bodies right now. Green points out that this transition period is, in a weird way, the best possible moment for AI to start getting seriously good at cryptanalysis.
The logic goes like this: new cryptographic problems haven't had decades of human scrutiny the way RSA has. If an AI system can probe them hard and either break them or fail to break them, that's genuinely useful information, not a catastrophe waiting to happen. Green's caveat is real, though. This only holds if AI doesn't blow through every hard problem we've built cryptography on simultaneously, and if we're not, as he puts it, living in Impagliazzo's Minicrypt — a theoretical world where one-way functions exist but the stronger cryptographic primitives we rely on don't.
Assuming neither of those grim scenarios plays out, Green's best-case outcome is oddly optimistic. AI systems chewing on these new post-quantum candidates could give cryptographers actual confidence that the underlying problems are as hard as everyone hopes, and the field of cryptanalysis itself could come out more rigorous for having been stress-tested by machines. He tacks on a plain 'hopefully' at the end, which is the kind of hedge that tells you even the experts aren't fully sure which way this goes.
My take — AI-written commentary, not fact-checked reporting
I find it refreshing when a cryptographer responds to 'AI might break encryption' with nuance instead of panic, because Green is essentially saying the real risk isn't AI cracking crypto, it's us not stress-testing our new standards hard enough before they ship. That's a much more useful worry, and it's the kind of measured take the AI-security discourse desperately needs more of instead of another round of doom or hype.
Read more about this at: Simon Willison's Weblog
