AI / OpenClaw Interview Questions
What is the difference between running OpenClaw directly on a host system versus inside a sandbox?
| Direct on host | Inside a sandbox |
| Agent runs with the full permissions of the launching user account | Agent's access is constrained to whatever the sandbox boundary allows |
| A compromise can reach files, SSH keys, and anything else that user can access | A compromise is contained within the sandbox in most cases |
| Simplest setup, lowest friction | Requires additional setup, such as Docker configuration |
| Reasonable trade-off on a fully trusted, isolated personal device | Recommended for shared infrastructure, exposed deployments, or less-trusted skills |
The practical guidance boils down to matching the isolation level to how much you actually trust everything the agent might execute, a fully trusted personal laptop with only vetted skills is a very different risk profile than a shared server running community skills from ClawHub.
What permissions does an agent get when running directly on the host?
When is running directly on the host a reasonable trade-off?
More Related questions...
