Matthew Green

We get a subexponential time algorithm for 3SAT and apparently the models couldn’t even find one little cryptanalysis result. I guess cryptographers have just been killing it!

I asked Astra to unredact a name/email that had been poorly redacted (one row of pixels on the top of one copy, one row showing on the bottom of another.) I put it into Codex and asked it to find the font, and brute-force the letters until it found a match. I didn’t expect much. It took 10 minutes.

Anyway it turns out that the answer is to combine frontier image models with abliterated text models, and write an unredaction tooling pipeline that just takes a stack of PDFs and hunts for bad redactions with only a little bit of (text) model help.

I asked Astra to unredact a name/email that had been poorly redacted (one row of pixels on the top of one copy, one row showing on the bottom of another.) I put it into Codex and asked it to find the font, and brute-force the letters until it found a match. I didn’t expect much. It took 10 minutes.

The only sticky part is that you have to lie to the models to get them to do this. They’ll ask you where the inputs came from, and you have to say it’s your own test data or else they refuse. I assume it’s possible to pipe this together with direct image calls and deterministic code, to avoid that.

Everyone over on X very excited that they saved the Internet from PHP vulnerabilities, and I’m thinking: if PHP is this important in 2026, we deserve whatever we get.

Why are the fridges still connected to the Internet.

I’m now running a little fleet of my own refrigerators.

Me: makes silly offhand comment about AI and public key cryptanalysis. People freak out. Also me: does a serious analysis of cryptanalytic effort so far set against recent LLM math results, which is (IMHO) wildly more terrifying. Nobody notices.

Everyone is very concerned about being respectable, so I’m going to be the goofball who raises worst-case possibilities. I think there is a 1% chance we live in Minicrypt, and a 15% chance we functionally lose confidence in our existing public-key encryption algorithms.

The problem here is that the speed of AI producing surprises, and the speed of human beings replacing standards (even with the very best AI assistance) are just orders of magnitude different. You only recover from a surprise like this if you do the preparation in advance.

Everyone is very concerned about being respectable, so I’m going to be the goofball who raises worst-case possibilities. I think there is a 1% chance we live in Minicrypt, and a 15% chance we functionally lose confidence in our existing public-key encryption algorithms.

These numbers are based on vibes, but nobody has better ones. There’s no upside to saying this. I’m not selling anything. I’m not going to get a prize for being right, and everyone will make fun of me for being wrong (even before I’m actually proven wrong ;)

If you haven’t asked abliterated GLM 5.3 to hack the Internet’s core routers, then you’re a happier person than I am right now.

Abliterated GLM 5.3 is like a nuclear weapon that’s fun to be with.

If you haven’t asked abliterated GLM 5.3 to hack the Internet’s core routers, then you’re a happier person than I am right now.

So I told GPT-6 to make me a virtualized simulation of a water treatment plant and now I’m spending a lot of time saying “please document that this is just for defensive purposes” because I realize how bad it looks.