From Friday Deep-Dive · Friday, October 23, 2026 · 6 min read
The AI lockdown checklist
AI assistants can now click, download and run things on your computer. That's useful, and it also means a bad website or a booby-trapped document can try to trick them. Here are the three rules I follow and the settings I use in each tool, written for people who are new to this.
Two resets in one week taught me this the hard way. Here's what happened, then the settings and habits I use now. You don't need to know anything about computers to do any of this. These apps change often, so if a button looks a little different on your screen, look for the closest match.
What happened to me
Last week I had an AI helper doing research for me on the web. I had turned off the setting that makes it ask me before it does anything. My best guess is that it landed on a sketchy website and picked something up there.
Claude later found a hidden program on my work computer called "RuntimeUtil." Every two hours, it quietly reached out to a bad website, downloaded new code and ran it. I can't prove exactly how it got there. The records were on that computer, and it won't turn back on.
Cleaning up took me more than 30 hours over three days. I changed every password, key and login I own. And I had kept my list of them in a spreadsheet and in an AI chat window. Don't do that.
Then on Wednesday, Oct. 21, Railway, a company that hosts websites and apps for me, said it had been breached. It told every customer to change every password and key stored there. So I did a big chunk of the work all over again.
One mistake was mine. One was a company's. I paid for both.
The three rules
These matter more than any setting.
- You say yes before anything risky happens. If the AI wants to download, install, send, buy or delete something, you click "yes" first. Skip anything called "auto," "always allow" or "don't ask again."
- Never give the AI your passwords. Don't type them into a chat. Don't keep them in a spreadsheet or document the AI can open. Use a password manager.
- Never follow a website's "press these keys and paste this" instructions. A real "I'm not a robot" check never asks you to press keys on your keyboard and paste something. That's a known scam. Close the page.
Why this matters now
AI helpers don't just answer questions anymore. Some can click buttons, fill in forms and visit websites for you. A sneaky web page, email or document can hide instructions meant for the AI, telling it to do something you never asked for. Anthropic and OpenAI, the companies behind Claude and ChatGPT, and Britain's national cybersecurity agency all say there's no complete fix for this yet. Your "yes" click is the safety net.
Lock down Claude in Chrome
Claude in Chrome is the Claude helper that opens in a panel beside your web pages and can click around websites for you.
- Make it ask first. In the Claude chat box, find the mode setting and pick Manually approve. Don't pick "Automatically approve" or "Skip all approvals."
- Approve one step at a time. When it asks to do something on a website, choose Allow this action, not "Always allow actions on this site."
- Give it its own Chrome. Set up a separate Chrome profile just for AI work. Don't sign in to your bank, email, health or government accounts in that profile.
- Clean up once a month. Click the Claude extension icon, then the three dots, then Extension settings, then "Your approved sites." Remove any you don't need.
Lock down Chrome itself
- Turn on the strongest protection. In Chrome, open Settings, then Privacy and security, then Security, and choose Enhanced protection.
- Check your add-ons. Click the three dots in the top right corner of Chrome, then Extensions, then Manage Extensions. Remove anything you don't recognize or don't use.
- Keep it updated. If Chrome shows an "Update" button in the top right corner, click it.
- Same rule for any built-in AI. If Chrome or any other app offers to let an AI act on websites for you, set it to ask before each step.
Lock down the Claude app
- Use incognito chats for anything private. Click the ghost icon at the top right of a new chat. Those chats aren't saved to your history or to Claude's memory. Anthropic still keeps them for 30 days, so don't paste anything you couldn't stand to have stored.
- Control what it remembers. Go to Settings, then Memory. Turn off "Generate memory from chats." Then choose Pause memory (keeps what's saved but stops using it) or Reset memory (erases it for good).
- If your office pays for Claude, ask whoever runs the account to set connected tools to "Needs approval."
Lock down ChatGPT
- Read before you click yes. When ChatGPT asks you to confirm something, read what it's about to do first.
- Stay at the screen while it works. If you ask ChatGPT to visit websites or do tasks for you, watch it. Don't walk away.
- Keep it logged out. Don't let it use your logins unless the job truly needs them.
- Type your own passwords. If a site needs a password, ChatGPT hands control back to you. Type it yourself. Never type a password into the chat.
- Give small, specific jobs. "Find the city council agenda for Nov. 3" is fine. "Go through my email and handle whatever comes up" is not.
- Nothing on autopilot. Don't set up scheduled tasks that can send email, post to your website or touch payment accounts. Nobody's there to say no.
- Control what it remembers. Go to Settings, then Personalization, then Memory. Turn off "Reference saved memories" and "Reference chat history" if you don't want them. Use Temporary Chat for anything sensitive.
- Keep your chats out of training. Go to Settings, then Data controls, and turn off "Improve the model for everyone."
Make a list of your important accounts
Do this before the next breach. When a company you use gets hacked, the hard part isn't changing passwords. It's remembering which ones you have.
Write down every important account: email, bank, credit cards, your website, social media, and any AI service with your credit card on file. For each one, note how to change the password. Keep the list in a password manager like 1Password or Bitwarden, not in a spreadsheet or a chat. That's where I went wrong.
Then:
- Add a second lock. For email and banking, turn on an authenticator app or passkeys. Once that works and you've saved your backup codes, turn off text-message codes where you can. CISA, the federal cybersecurity agency, warns that text-message codes are easier for scammers to get around.
- Set a spending limit on any AI or payment account, so a thief can't run up a big bill.
If something goes wrong
- Stop. Turn off Wi-Fi or unplug the computer from the internet. Don't keep using it.
- Change your passwords from a different device, like your phone or another computer you trust. Start with email and banking.
- Check for strange activity: your bank and credit card statements, and your email settings for any rule that forwards your mail somewhere else.
- Have a professional wipe and reinstall the computer. Some bad programs hide and come back after a normal cleanup. Mine ran every two hours, and I never saw it.
- Write down what you remember: dates, what you clicked, screenshots. That helps someone figure out how it got in.
If you write code
If you use AI coding tools like Claude Code or Codex, hit reply and I'll send you the technical version of this checklist.
These are MY suggestions and things I did
Do you own research if you are not comfortable with your settings. If anything, ask in the chat of the AI tool you use what to do.
Try this and tell me what happened.
Did it work? Hit reply. Did it come back sounding like a stranger? Hit reply harder. That’s the interesting failure, and it’s almost always fixable.
Get the next one in your inbox
Tuesday brief + Friday deep-dive. Free. Unsubscribe anytime.
Subscribe — free, twice weekly← Back to What a week, my AI horror story from the past seven days