← Back to the A-Z glossary
AI glossary · P1

Sandbox

A sandbox isolates the agent to limit the damage if an action goes wrong.

Security 4 min read Updated 2026-05-22
— Definition

Sandbox, in plain words

A sandbox isolates the agent to limit the damage if an action goes wrong.

Present the sandbox as a cage that limits damage, not as a prison.

A concrete example

Codex runs in an isolated container. If the agent goes off the rails, your machine stays intact and the container gets tossed.

Why it matters

Without a sandbox, an agent that slips up can wreck your machine. With one, the worst case is a reset.

You see it in Codex cloud, agents that run code, tests, and controlled environments.

Don't mix it up with

Tool permission: A tool permission decides what an agent is allowed to read, write, or run.

Least privilege: Least privilege means giving the agent only the rights it actually needs.

Common mistakes

  • Giving the sandbox access to the host system.
  • Forgetting to cut off internet in the sandbox when it isn't needed.
  • Believing a sandbox equals total security.

Quick checklist

  • First I check whether the word names a concept, a tool, a risk, or a metric.
  • I tie it to a concrete case: Codex runs in an isolated container. If the agent goes off the rails, your machine stays intact and the container gets tossed.
  • I keep the main trap in mind: Giving the sandbox access to the host system.

Quick questions

What is Sandbox in AI?

A sandbox isolates the agent to limit the damage if an action goes wrong.

Where will I run into Sandbox?

You see it in Codex cloud, agents that run code, tests, and controlled environments.

Which word should I read next?

Start with Tool permission, Least privilege, Guardrail.

Want to keep going in order?

Head back to the full glossary, search a word, then open only the pages that deserve more than a short definition.

Open the AI glossary