Agents use tooling calls to run commands on their hosts.
policy based sandbox (e.g. nvidia openshell)
- uses policies to enforce rules on the level of system calls, file access and network calls
- enforcdement happens outside the agent
- outbound network requests go through a supervisor outside the sandbox, checking the policies
- possibly on separate hardware such as nvidia sentry, holding tamper evidence outside of agent hosts
network sandbox
To avoid agents accessing local networked devices, they are typically set up in a separated, secure environment such as a VLAN or virtualised network: Setup Openclaw on Restricted Proxmox