Is it safe to let AI control your computer?

It can be kept under control, but it is not free of risk. Hey Atu gives an AI app real access to your computer: commands run with your user account’s permissions, and there is no sandbox. This page says how Hey Atu limits that access, what our relay can see, what it cannot protect, and how to report a problem.

Get started

By default: Off. No app can connect until you allow AI apps
Access: Each AI app needs your approval on the computer, with a code
Network: The computer connects out. No inbound port
Stop all: heyatu mcp off ends every AI app’s access

Last updated 24 October 2026

What a connection can do

An AI app needs your approval. Google sign-in alone gives no access. The first time an app connects, a window on the computer names the app and shows a code. You click Allow.
It can do what your user account can do. It reads and writes the files you can reach and runs commands. There is no sandbox.
AI apps cannot use your model sign-ins. They use their own models. Only the eight tools on your computer do the work.
Your computer keeps a list of calls: the tool, the time, the result and a short preview of the input. It does not keep the full content.
A paired phone has full control. This applies to the Hey Atu iPhone app, which testers use today. A pairing code works once and only for 120 seconds. Scan only a code on your own computer.

What the relay can see

AI apps: not end-to-end encrypted. ChatGPT and Claude call from their own cloud, so their servers must read the results. Our relay forwards this traffic. It does not store or log it.
Phone and computer: end-to-end encrypted. Chats in the Hey Atu app are sealed. The relay forwards them and has no key.
No inbound port. Your computer connects out to the relay. You do not open a port or use a VPN.

Risks that remain

Prompt injection. Text in a file or on a web page can tell an AI to run a command. Keep “ask before use” on for tools that run commands or change files, where your AI app has this setting.
Your AI app account. A person who controls your ChatGPT or Claude account can use your connected computer. Turn on 2-step verification there.
Mistakes. A model can delete or change the wrong file. Keep backups of what matters.

What you can do

Allow only AI apps that you trust, and check the code before you click Allow.
Stop all AI apps at once with heyatu mcp off; turn them on again with heyatu mcp on.
Disconnect Hey Atu in an AI app that you no longer use.
Do not share a pairing code or a photo of one. It gives control to the person who scans it.
Lost phone? Remove it on your computer: heyatu phones, then heyatu forget.

Compared with self-hosted agents

With an agent that you host yourself, such as OpenClaw, much of the risk comes from its setup: a gateway open to the internet, or a third-party skill with malicious code. Hey Atu has no inbound port and no skill store, but the risks above apply to it too. See Hey Atu vs OpenClaw.

Report a vulnerability

Email security@heyatu.com. Tell us what you found and how to reproduce it. We aim to reply within 7 days.

Please test only against your own computers and accounts. Do not access other people’s data, and do not disrupt the service. If you follow these rules and give us time to fix the problem before you publish it, we will not take legal action against you.

Know the limits first. Then decide.

Set up in a few minutes. 100 free actions, no card.