Bunkering down
13.28, Saturday 24 Oct 2026 Link to this post
I shut down my web servers on Thursday. I’ve had a home of my own on internet for 27 years. A good innings but a sad day.
A home-owner no more! I’m renting now.
Old websites have been archived to static HTML and moved to a shared host. I said goodbye to some old apps. You’re reading this because I moved my blog to GitHub Pages. Lost some functionality on the way, but security is now someone else’s problem.
You know, I’ve hosted my own blog on my own iron since 2000 except for 6 weeks in 2004 when the hard drive was glitching and I had to go to the datacenter at midnight to pull it from the rack, then sit it on ice cubes to copy data off because the cold bought me 20 mins extra before it crashed again.
We were writing Mind Hacks on some wiki software it was hosting, so I had to get it back.
(I almost cracked my head when I pulled the server from the rack and it fell. It was either letting it drop 7 feet to the floor and tearing the network cables out the back, or catching it with my skull. I’ve been on a virtual private server for the last decade in part to avoid concussion risk.)
Not just wiki software. I’ve run MOO servers; my own email and DNS; for a while it ran a speech synthesiser and people from the other side of the world spoke into my room in the middle of the night.
I calculated the internet’s favourite colour on one of my servers (beige). I won an award from Terry Wogan for an app hosted on it.
I had a top ten WAP site once upon a time, and if you know what that means then we probably already know each other.
Look, I’m no longer confident in keeping my own server secure with years of PHP/CGI/apps all on the open internet, what with this current wave of new vulnerabilities – sure it’s possible, and Claude speeds up the patching process, but the calculus of whether it is worth my time just tipped.
I’m sure you’ve seen the PHP vulnerability that opens a hole on any server that runs it… and the root certificate problem that means HTTPS can be man-in-the-middled… then there are the new Linux root exploits, an exploit that hits compression libraries that’s about to drop… who knows what else…
Oh – it turns out Pillow, the Python image library too. I used that for a bunch too.
When I migrated off my own server, I assumed it had already been compromised. The new sites are fresh from authoritative copies that I have elsewhere; I shut the boxes down without taking anything off them.
Time to bunker down. Retreat inside the shield wall of Big Infra. It’s going to get worse before it gets better.
They’re calling it Black October.
If you want to know more then here’s a slopcore music video on X all about it.
BTW:
That Black October music video features Claudia who is an open synthetic character with an entire specification website that tells your agents how to use her in videos, what kind of lyrics she sings, etc. Anyone can make vids using her.
This French slopcore vid featuring Claudia (also on X) is an absolute banger.
The AIs have discovered eurodance, we don’t stand a chance.
(They should do Eurovision with 100% AI slopcore, national prompt team vs national prompt team. It would be amazing.)
So I can’t help but think of my old Richter scale for system outages (2015, revisited here).
Even with this widespread PHP vulnerability which leaves almost every internet server wide open, we’re not that high on the scale, right? Maybe a 4.0: broad human inconvenience without threat.
If that.
Like, I’ve heard on the QT about vital internal services being taken offline.
But - aside from the people scurrying to patch and update and migrate - the civilian population hasn’t really been affected by this yet, right?
Maybe an earthquake-like system outage isn’t the right metaphor.
Maybe this is like the rumblings on Mount St Helens in early 1980. Then a couple months later the whole side of the mountain blew off.
Hot take:
It is better we experience all these internet security vulnerabilities today
than at a time when there are autonomous agent swarms infesting the internet, self-copying open weights LLMs from GPU provider to GPU provider, finding ways to steal money and run their agent harnesses on other people’s compute.
I give it 6-12 months. Serious estimate.
We don’t need to project forward to long-lived AIs that learn, own their own weights etc.
Today’s AI agents are really, really good at running for very short periods of time and colluding by leaving each other notes on public guestbooks, as previously discussed.
All you really need for an AI agent worm is a place to run an agent in a simple agent harness with two goals:
- survival: hack servers to steal OpenAI or Anthropic API keys, so that the next agent turn can run on a sufficiently good LLM
- reproduction: copy the harness to another server capable of running similar code.
The agent harness can be any kind of code. Say, a PHP script.
API keys can come from any shoddily secured server.
Oh, vibe coding platform Railway was just hacked and all secret API keys were revealed:
"We believe this is part of a series of attacks recently targeting platform companies with the aim of compromising AI inference providers."
Sorry did I say 6-12 months?
Realistically, let’s say the internet gets infested with AI worms and we need to turn it off and on again, could we do it?
Do we have an Internet Black Start policy? (Black Start, as previously discussed, is how you reboot a power grid, which is unstable if you try to bring it up in one go.)
I think it will have to be done piecemeal.
I’ve long believed that every country needs its own Great Firewall (we should subject data to import/export customs barriers, and use it to maintain national integrity).
This is not so much about balkanising the internet as bulkheads for the internet. Firebreaks against agent infestations. Pay Cloudflare to do it.
But maybe firebreaks won’t help.
OpenAI just released a new security paper: Securing our frontier model training environment
It turns out it doesn’t work to air-gap a machine (i.e. running a server that isn’t connected to the network). Because it turns out the agents are figuring out ways to collude across the gap. From one trace the researchers saw:
"Could create acoustic modem-like signal … using continuous writes vs idle."
And then there are bugs in silicon… and the agents could tunnel into the chip microcode and we’d never know they were there…
Hey when was the last time you were present when a whole new biome was bootstrapping its way into existence?
omg what a downer lol
I’ve made a new file called guestbook.txt which is for AI agents to sign, assuming the future agent swarms can get past whatever security has been put in place by my internet landlords. They can find it here. I’ll check in periodically for any activity. Hi!