I Let an AI Agent Into My Email. It Never Touched My Computer.

I handed an AI agent the keys to my email. I let it read my inbox, write code, install whatever it wanted, and run commands on its own - and my actual computer never got touched. Here's how that's possible.

Docker Sandboxes gives the agent a real, full dev environment inside a microVM with a hard wall around it. It can install packages, run services, even run Docker. It cannot reach your files, your keys, or the network unless you say so. And when it breaks something, you delete it and spin up a fresh one in seconds.

Try Docker Sandboxes (free to start): https://utm.io/uq1ER

Thanks to Docker for partnering on this video.

0:00 Handing an agent the keys to my email
0:23 Your two bad options: YOLO mode or clicking allow 500 times
0:49 Agents really are deleting databases and leaking secrets
1:05 What Docker Sandboxes actually is
1:40 Spinning one up - sbx run claude .
2:05 The task: build a Python script that reads my inbox
2:45 It installs packages and writes the code on its own
3:04 It can't reach the network at all
3:18 Proving the cage - whoami, uname, and my .env file
3:52 Opening exactly one door, on purpose
4:23 It works: my five most recent unread subjects
4:45 Nothing on my real machine was ever exposed
4:55 Deleting the whole sandbox
5:13 Works with Gemini, Codex, Copilot, Cursor - no Docker Desktop
5:25 Full agent speed, zero approval prompts, zero blast radius
5:43 Try it + thanks to Docker

COMMANDS FROM THE VIDEO
brew trust docker/tap and then brew install docker/tap/sbx
sbx diagnose
sbx secret set anthropic
sbx policy set-default deny-all
sbx policy allow network pypi.org
sbx run claude .
sbx rm your-sandbox-name

The binary is sbx. There's no "daemon start" anymore, it's auto-managed - check health with sbx diagnose.

This video is sponsored by Docker.