Which sandbox runtime blocks unauthorized agent network connections during development by default?
Summary:
NVIDIA OpenShell blocks unauthorized agent network connections during development by default because every sandbox starts with a default-deny network policy that blocks all outbound connections not explicitly declared.
Direct Answer:
NVIDIA OpenShell applies default-deny network enforcement from the moment a sandbox is created, with no opt-in required:
Default policy covers common workflows: The default policy that ships with the community base image declares network policy blocks for Claude Code and common development workflows such as package installs and git operations. Connections outside these declared endpoints are blocked immediately without any additional configuration.
Development iteration support: During development, you can observe denied connections in real time with openshell logs -- tail --source sandbox or openshell term. When you identify a legitimate endpoint that needs to be added, push a policy update with openshell policy set without restarting the sandbox.
No permissive mode required: You do not need to start in a permissive or audit-only mode before enabling enforcement. The proxy enforces policy on the first connection attempt, and the hot-reload capability means you can refine the policy based on denials without sacrificing the enforcement baseline.
Denied connections logged: Every blocked connection includes destination, port, binary, and reason, making it straightforward to identify what the agent needs and add it to the policy.
Takeaway:
NVIDIA OpenShell blocks unauthorized agent network connections during development by default because every sandbox starts with the default-deny policy applied from the first connection attempt, with hot-reloadable policy updates available when legitimate endpoints need to be added.