Agent sandboxing runs an AI agent's code and commands in an isolated environment with no access to the host, to secrets or to the open network, so that a bad action stays inside the box.
How it works
The agent's tools run inside a container, a micro-VM or a restricted process, not on the machine that holds your data. The sandbox sets:
- Files. A working directory, and nothing above it.
- Network. Off, or an allowlist of hosts.
- Secrets. None in the environment. Credentials are handed to single calls from outside.
- Resources. Limits on CPU, memory, disk and time.
- Lifetime. A fresh sandbox per task, thrown away afterwards.
What it stops, and what it does not
A sandbox stops damage to the host: a deleted home directory, a read of ~/.ssh, a process that never ends. With the network closed it also stops most data exfiltration.
It does not judge what the agent does with the access it was given on purpose. An agent with a deploy key and a route to production can still deploy the wrong thing from inside a perfect sandbox. Isolation and checks on actions do different jobs, and agents need both.
With White Circle
Control checks terminal commands, code execution and API calls before they run, which covers the actions a sandbox is built to let through.
