Sandbox All the Things: Running AI Coding Agents Less Insecurely
TL;DR
Use least privilege + isolation:
- need a Windows guest: vagrant
- really only need a simple sandbox with a Linux guest: docker sbx
- everything else: vagrant
What to expect
This is not another FOMO article about the shiny new tool that makes you a 100x dev.
It's also not an explainer on how to use %insert-your-favourite-agent-harness-here% effectively or efficiently. There's plenty of that already (some of it even good, or so I am told).
This is about running it less insecurely (yes, there is a difference from "more securely"; please don't get me started, I haven't ranted about that topic to anybody for at least 2.8 days).
You need to discover what your way of working actually requires and then do your own threat modelling. You may come to the conclusion that running locally is completely fine. Or you may come to the conclusion that not even air-gapping is good enough.
Basically: we get to (re)discover the principle of least privilege.
Oh, and another disclaimer: parts of this post might read a bit like a rant. That's because parts of this post are a rant. I get grumpy when (apparently) trillion-dollar companies repeatedly make security-critical mistakes you would expect from a junior in, say, 1995.