A field guide to sandboxes for AI
Years ago I first learned this lesson doing basic malware analysis. The mental model was blunt but effective: run hostile code in something you can delete. If it breaks out, you nuke it.