deploy-sandbox

To deploy an AI agent with filesystem isolation and network egress control, use this skill to launch NAT agents as isolated OpenShell sandboxes. It applies Landlock filesystem protection and a default-deny network policy, brokering model traffic through a secure gateway to prevent unauthorized external access.

Landlock filesystem protection is a security mechanism used to create strict isolation boundaries for deployed NAT agents. It ensures agents operate within a governed sandbox environment, preventing unauthorized file access while allowing secure execution through a gateway-managed inference route.

Yes, deploying a governed sandbox requires a pre-built agent image, a configured NeMo Platform, and an OpenShell gateway. These components are necessary to enforce zero-egress governance and broker model calls through the secure inference route.

A default-deny network policy blocks all outbound traffic from isolated NAT agents by default. Model calls are brokered through a secure gateway-managed inference route, ensuring agents only communicate with authorized endpoints and preventing unauthorized external network access.

Yes, you can use the openshell-local executor to deploy sensitive agents on a local platform. It enforces strict security boundaries, including Landlock filesystem isolation and zero-egress governance, to ensure agents only access authorized inference endpoints.

The primary limitation of using Landlock and default-deny policies is that agents cannot access external networks or unauthorized files outside their sandbox. All model traffic must be routed through a pre-configured gateway, requiring a fully set up NeMo Platform and OpenShell gateway.