Jargon:Sandbox
1 min read
Say this at the party
“A walled-off space where an AI can run code and make a mess without touching anything you care about.”
In plain English
Think of a children's sandpit. Whatever gets built or knocked down stays inside the sandpit.
A computer sandbox works the same way. The AI gets its own limited space, with only the files and internet access it needs for the job. If it deletes the wrong thing or gets tricked, the damage stays inside.
It's what makes it safe to let an agent work without asking you before every step.
Where it came from
Security people have used sandboxes since the 1990s to run programs they don't trust. Agents made it everyday advice: Anthropic added sandboxing to Claude Code in 2025, and Simon Willison's rule is to run any unsupervised agent in a sandbox or on a throwaway machine.
See it in action
You let a coding agent run on its own overnight. It can read and change the project folder and download what it needs, but nothing else: not your photos, not your passwords, not your bank.
What people get wrong
- A sandbox doesn't make the AI's work right. It just limits the damage when it's wrong.
- Giving a sandboxed agent your real logins punches a hole in the wall. Use test accounts with a spending cap.
Go deeper
- ArticleAnthropic, 2025Making Claude Code safer with sandboxingHow a sandbox lets an agent ask permission far less often, safely.
- ArticleSimon Willison, 2025Designing agentic loopsPlain advice on where to let an agent run free.
- GuideWikipediaSandbox (computer security)The longer history, from before AI.
Where it comes up
Related jargon
Drafted with AI, checked by the AI Council (Claude, ChatGPT and Gemini) and signed off by Phin.