Matthew Green

AI propaganda is here, and this is only the beginning. Exactly as I predict in my book. A Chinese (apparent) subcontractor, is reportedly using AI combined with massive data collection and analysis to run targeted influence operations. www.nytimes.com/2025/08/06/u...

That’s exactly what they want. Project 2025 explicitly calls for shutting down foreign social media like WeChat and TikTok. (Trump’s implementation is more of an attempt to force their sale than a shutdown.) There is a strategic goal aimed at political stability; it’ll just be their politics.

Project 2025: What a second Trump term could mean for media and technology policies | Brookings Roxana Muenster analyzes Project 2025’s plans for technology, media, and communications policies and the implications for existing policies.

The Heritage Foundation does hate LGBTQ. But they also have a document that talks about amassing power and crushing every opposing power center in the United States. These two things aren’t mutually exclusive. www.brookings.edu/articles/pro...

I think there’s a vanguard of people who really just hate porn. I think the reason they’re getting so much deep support is because other more thoughtful people behind them see the implications of an unchecked Internet as bad, and view this as one (legislative and technical) tool to end it.

so the stupid canadian age verification bill is back in parliament so call your local MPs and tell them it’s bad and why it’s bad and don’t be ashamed to pander to them based on if they’re a tory or grit or otherwise - if you frame it in a way that mirrors their ideology, they might remember better

I admit it sounds a little looney tunes when I write it down that way and yet, we live in a looney tunes time.

For the ninth time I’m going to repeat my personal conspiracy theory that every democracy is fighting a battle to control its Internet against foreign influence, and “age verification” is not being pushed so hard because kids, but is actually one weapon being deployed for that fight.

It Looks Like a School Bathroom Smoke Detector. A Teen Hacker Showed It Could Be an Audio Bug A pair of hackers found that a vape detector often found in high school bathrooms contained microphones—and security weaknesses that could allow someone to turn it into a secret listening device.

The Halo 3C is a smoke/vape detector that Motorola sells for use in school bathrooms. It also has microphones inside.

A teen hacker found them at his school, and with another security researcher has now shown they could be hacked for audio surveillance.

www.wired.com/story/school...

Just sitting here with an unfinished algorithm for solving MLWE all I need is that one missing piece.

OpenAI continues to deny me GPT-5 presumably because it knows I’ve got a list of prompts that will break all cryptography if it’s 8% as good as Sam Altman says.

I’m saying this because I had to switch back from book writing (fun! easy! still harder than many things I’ve done) to research paper writing, and it’s so much slower and more intricate. And I’m not even the lead author doing the hard stuff.

And a lot of the time the difference between “good” and “bad” science is the kind of laziness that’s just Tuesday in software development.

You’ve got a field that basically involves obsessive people doing obsessive work for relatively terrible pay and then you’re like “I, as an expert in JavaScript development, demand that every result be replicated.” I mean ok that would be cool, but how and with what resources.

Anyway, I just want to be clear that I’m not patting anyone on the back. I’m just saying that this is absurdly more work than most other things people could be doing, for such a relatively small result. I think about this when people criticize science (often rightly) for lack of replication etc.

One of the things that amazes me about scientific research is how much work goes into it. It’s just crazy. I mean all things require work, but it’s just so far above anything else (writing, software development.) I understand all the complaints about replication, what’s missing is the labor.

Jury deadlocks on money laundering charge against founder of crypto 'mixer' Tornado Cash A U.S. jury deadlocked on Wednesday on money laundering and sanctions evasion charges against the founder of Tornado Cash, a firm that makes cryptocurrency transactions harder to track.

Jury deadlocks on money laundering charge against founder of crypto 'mixer' Tornado Cash reut.rs/45t4KVJ

Yes we did some work on this. It’s great, but then you need to make sure the LLM cover traffic isn’t weird. meteorfrom.space

Yes but it’s really easy to have AI generate stuff that “looks good” if you’re using the wrong definition of “looks good”, but might be really easily detectable. I guess maybe the real recommendation is that defenders need to think like censors, too.

I agree that scale is going to be an issue for them, I don’t think you’ll have a transformer processing every packet. I do think you could build an architecture that’s heavily adaptive and uses filtering rules developed by gen AI.

"How the Great Firewall of China detects and blocks fully encrypted traffic." is IMO the first glimpse into what this looks like, though.

The thresholds in Ex1 smell like something written by a decision tree and then adapted by humans.

Yes. It illustrates that (1) they care, (2) they’re happy to automate, (3) they’re operating at like .001% of the potential capability they could implement.

It also doesn’t help that DOGE cancelled a bunch of NSF grants working on censorship-evasion tech because they contained the keyword “censorship.”

I don’t have anything wise to say about this. I just believe there’s a huge amount of room to lock down national Internet connections in ways that are going to be hard to bypass. It’ll come as a surprise when it happens and it would be better if we were thinking about it now rather than later.

A lot of the tools we have for evading censorship work on the principle that nobody’s looking too closely, and when they get clever and use steganography it’s only a little bit clever; the bar is very low to sneak past the tools that exist today.

And maybe it will stay that way for “business reasons” but technically it’s due for a massive upgrade thanks to the new capabilities AI will be able to throw at network traffic in the next few years.

This is a field that’s been working for years to find clever ways to sneak traffic across China’s Great Firewall or out of Iran, and I think it’s become accustomed to a very lazy and mostly human adversary.

So I don’t like to generalize big conclusions from a few random conversations, but I feel like the censorship-evasion field of research is about to get clobbered by AI.

How I feel every time I visit X.

To give you a sense of how bad the GOPs current polls must be. Indiana is already heavily gerrymandered and there are two Democrats out of 9 seats.

“I gotta be honest, the whole thing is pretty nuts,” says JHU CS’ @matthewdgreen.bsky.social. “My best hypothesis is the phones are seeing a Wi-Fi access point that is slightly off from where they are.”