Sandbox

How it works

A sandbox defines in advance which folders a program can see, which networks it can connect to, and which commands it can run. If the program tries to go beyond this scope, the operating system or execution tool blocks it or asks a human for permission. Some sandboxes use a fully separate computing environment such as a container or virtual machine, while others set up a lightweight boundary using only the operating system's permission features. When the work is done, the whole sandbox can be deleted so no traces remain.

Why it matters

Today's AI agents do more than write text: they run code, edit files and access the internet directly. If an AI makes a mistake or is fooled by prompt injection (an attack that sneaks in malicious instructions), it could delete important files or send information such as passwords outside. A sandbox ensures that even if this happens, the damage stays inside the boundary. That lets people hand more work to agents without checking every step themselves.

Things to know

Having a sandbox does not make things completely safe. If the allowed scope is too broad, the boundary is pointless; if it is too narrow, the agent cannot do much. So sandboxes are usually combined with other measures, such as requiring human approval for risky actions (human-in-the-loop) and guardrails.

Example

Sandboxes are especially important for coding agents, because they run commands directly on the user's computer. As in the news covered on this site, 'GitHub, Copilot 로컬 샌드박싱 정식 출시,' developer tool companies are adding sandbox features so that AI coding assistants operate only within a set scope, even on a developer's own PC.

Where it came up on this site

• briefing GitHub makes Copilot local sandboxing generally available, also releases CLI local model discovery and a dedicated model for detecting leaked secrets 2026-10-07

Related terms

AI agent Coding agent Human-in-the-loop Prompt injection Guardrails Computer use