Sourced from: GKE Agent Sandbox docs, kubernetes-sigs/agent-sandbox
So here is the thing. We are asking AI agents to do increasingly wild stuff. Write code. Browse the web. Run shell commands. Spin up browsers. Click around. Deploy websites. Use a computer like a human would. And for the most part, they are doing it.
But there is a lurking problem that keeps platform engineers up at night.
Where the hell is this code running?
When you ask an agent to "run this script" or "install this package" or "check if that endpoint responds", you are effectively handing over a loaded gun to a very enthusiastic intern who has read every programming book but has absolutely zero survival instincts. That code could be malicious. Or buggy. Or it could just go nuclear on your host system because the agent decided to rm -rf / thinking it was cleaning up temp files.
