Sandboxes for Code-Running Agents

Agent Safety Advanced 4 min read Published
In this lesson

Data-analysis agents, coding agents and “code interpreter” tools all run code the model wrote. That code might be buggy (an infinite loop), wasteful (filling the disk) or, after a prompt injection, hostile (reading your keys). Never run it in your main program. Run it in a sandbox: a place where the worst it can do is harmless.

🎬 The playpen

🔒 This is a premium lesson

It is part of the Agentic AI Engineer career track.

🎬 Live animations · 🧪 Tested code · 💼 Projects · 🎓 Certificate

See the track

Already a member? Sign in

Progress is saved in this browser. Create a free account to keep it on every device.