Is it safe to let AI control your computer?
What can go wrong when AI clicks for you, and the five safeguards that make it safe: plans, tiers, a check outside the AI, live view and a log.
October 31, 2026
Letting AI click on your computer sounds risky, and it should. A tool that can press any button can press the wrong one. The question isn’t whether to trust the AI. It’s what stands between the AI and your computer.
What can actually go wrong
- A misunderstanding. You said "send it to Dana" and there are two Danas.
- A wrong click. The agent picks the button next to the one it meant.
- A trick. An email or web page contains text written to give the AI orders. This is called prompt injection; we explain it in When a web page tries to give your agent orders.
- Over-reach. It does more than you asked, because it thought it was helping.
Five safeguards that matter
1. A plan before anything happens
2. Every action is sorted
Before any click reaches the screen, it’s sorted into read (look at a window), change (click, type, move files) or sensitive (send, pay, delete, install, sign in, press Enter in a form, read the clipboard, take a screenshot). Sensitive steps always ask.
3. The check lives outside the AI
The sorting is plain code that looks at what’s really about to happen: which app, which button, which words are on it. The AI can’t lower a tier, and nothing it reads can. That’s the difference between "the AI promised to ask" and "it can’t not ask".
4. You watch it live
An orange hand moves on your screen. You see every step as it happens and can stop it at any point.
5. Everything is logged
Every action, who asked for it, and whether you approved it is saved on your computer.
What stays locked
So, is it safe?
It’s as safe as its brakes. An agent with no plan, no tiers and no outside check is a risk. One that shows you the plan, stops before anything that can’t be undone, and can be stopped any time is something you can hand real chores to. Read the details on our Safety page.