Two weeks, one network
About eight years ago I set up the network in our house. I ran a cable to every room, put the servers in a rack in the basement, and gave each machine one job. One old server did a little of everything: it was the router and firewall, the mail server and the web server for my domain. It worked. It kept working. And because it kept working, I stopped paying attention to it.
That's the trap with a system that never breaks. The operating system on that server had stopped getting security updates years ago, and it was the one machine facing the internet directly.
So a few weeks ago I sat down with Claude and said, more or less, "let's fix this." Here's what happened over roughly two weeks.
What we did
- Moved mail and the website to a newer server running a current version of Fedora Linux. Mail now uses encrypted connections and signs outgoing messages, so forged mail from my domain is easier to spot.
- Moved DNS to Cloudflare, with a small script that updates my records automatically if my internet provider ever changes my address.
- Set up offsite backups for every machine in the house, nightly and encrypted, with deletion protection. (That one gets its own post.)
- Built a new router on a small dedicated firewall box, so the internet-facing machine only does router work, runs current software and gets updated regularly.
- Tidied up the edges: password logins from the internet turned off in favour of keys, a web admin console nobody needed switched off, and the Wi-Fi access point updated and given a new admin password.
- Brought everything up to date: every machine is now on a supported operating system.
Along the way we wrote one long plan document that recorded every step, every decision and every check. It turned out to be the most valuable thing we made. When I come back to this in a year, it will tell me why things are the way they are.
What it was like
I expected a clever search engine. What I got was closer to a very fast, very patient colleague who can actually run the commands. I'd describe a goal, Claude would look at the real machines, propose a plan, and do the work once I said yes. It checked its own work and reported the results, including the ones that didn't go as planned.
A few things stood out:
- It asked before changing things. Anything that touched a running system, or that I couldn't easily undo, waited for my OK. Read-only checks it just did.
- It respected my decisions. When I said I didn't want something, such as a password lock on a desktop machine, it stopped suggesting it.
- It kept secrets out of sight. It never printed a password or key, even its own. It checked them by length or by whether they worked.
- It knew when it needed my hands. Patching cables, finding a BIOS setting, clicking through a web page for my DNS provider: it walked me through those step by step, and adjusted when the screen didn't look the way it expected.
Did it make the network safer?
I asked Claude for an honest guess. Its answer: with the new router in place, the network is several times safer than before, perhaps 80 to 90 percent less exposed to the kinds of attacks that actually hit home networks. The biggest single gains were retiring the out-of-date internet-facing machine and having real, tested backups.
That's a judgment, not a measurement, and Claude said so. But I've done this kind of work for a long time, and it matches my own sense of where the risks were.
What I'd tell a friend
- Give it real work. The value showed up on the boring, important jobs I'd been avoiding.
- Stay in the loop. I approved every change and asked "are you sure?" often. It was wrong sometimes, as I describe in Mistakes AI made.
- Keep a written record. Ask it to update a document as you go. That record is worth as much as the work itself.