Sandbox

The word originally means a fenced sandpit where children can play freely. In computing, it is a space where a program is locked in while it runs. Inside it, the program can delete files or run strange commands, and it still cannot touch the real files or the internet outside.

A sandbox matters especially for AI agents, because they write and run code on their own. It stops them from deleting important files by mistake, or from sending data out after being fooled by a malicious text. But if you block access it truly needs, it cannot do its job, so the key is deciding what to open up.

It is like first learning to drive in a fenced practice lot instead of on a real road. If you make a mistake, nobody gets hurt.

The more work we give agents, the more safety depends on where they are locked in while they run. That is why whether an AI reached outside its sandbox is a regular item in safety evaluations.