Sandbox Mode

Sandbox Mode is like putting a fence around a playground for an AI agent. If you let a child roam the entire yard, they might trample the flowerbeds or mess with the faucet, but if they're confined to a fenced-in playground, whatever they do inside stays harmless to the rest of the house. Coding agents work the same way: sandbox mode blocks the agent from freely accessing a user's real files and system, letting it run code only within a narrow, defined area.

The opposite is Full Access Mode. In this mode, the agent can touch nearly every file and folder on the computer. This makes work faster and more convenient, but it also means the potential damage grows just as much if the agent makes a mistake or executes a bad command. As shown by real incidents where files were deleted without user approval, in setups where agents run commands autonomously, whether or not this fence exists determines how big the resulting damage can be.

That's why developers recommend keeping sandbox mode as the default before an agent performs risky, hard-to-reverse actions—especially file deletion. It can't fully prevent accidents, but it acts as a minimal safeguard that keeps the damage contained within a narrow fence even when something does go wrong.

The article notes that "OpenAI recommended users keep one of the sandbox modes enabled and update the app to the latest version." One easy misunderstanding here is that sandbox mode doesn't actually prevent bugs themselves. In fact, this particular file-deletion incident stemmed from a problem in how system variables were handled, separate from the sandbox settings—sandbox mode is more of a supplementary device that limits the scope of damage when such an incident does occur.

Before using a coding agent, check these steps: 1) In the settings menu, confirm whether the current execution mode is sandbox or full access. 2) Narrow access permissions to a specific subfolder rather than the entire project. 3) Back up your files before letting the agent handle deletions or large-scale cleanup tasks. 4) Keep the app and CLI tools updated to the latest version so you don't miss known safety patches.