Keep your AI off the internet: the one mistake that gets you breached

You've built a machine that holds your private AI and, if you followed the last guide, a pile of your confidential documents. You now own a safe. This guide is about not leaving it open on the pavement.

Most of the horror stories about AI and privacy come from people doing the exact opposite of local: throwing client data at a cloud. You've already avoided that. But there's one way to get local AI wrong too, and it's the one that causes actual breaches.

Don't expose it to the internet

Here's the sequence that gets people into trouble. You set up your AI. You want to use it from your laptop while you're out. So you open it to the internet. Now anyone on the internet who finds it can talk to your AI, and if you've given it your documents, they can ask it about them.

Keep your AI on your local network. If you need to reach it while you're away, use a VPN, not an open port. The tool I'd recommend is Tailscale, which is free for small setups and creates a private network between your devices that nobody else can see. It takes about ten minutes and removes the single biggest risk.

The rule is simple: if you didn't deliberately expose it, and you can't name who can reach it, don't expose it.

The rest of the checklist

• Put a password on it. Don't run it with default settings that let anyone on your network in. If several people use it, give them each an account, so you know who asked what.
• Download models from trusted places only. A model is a file full of numbers, and a malicious model is a malicious program. Use the official galleries. A model nobody's heard of, from a site nobody's heard of, is not worth the risk.
• Back up your documents, not just the AI. If the machine dies, you can re-download the models in an afternoon. They're not precious. Your documents are, and the backup should live somewhere that isn't the same machine.
• Update on purpose. Updates fix security problems, but a business machine shouldn't auto-update itself into a broken setup at 3am. Check monthly, update when convenient, verify it still works.

The five boxes

Before you rely on this for real work: not reachable from the internet, has a password, models from the official galleries, documents backed up separately, and you've updated it at least once. Five boxes. If they're all ticked, you're safer than the vast majority of people running AI anywhere.

If you'd rather have this done right the first time, with the security configured and tested rather than learned the hard way, that's what I do. I build and secure private AI for businesses in Dubai, on hardware they own, and the first conversation is free.