Why it matters
NVIDIA’s AI Red Team provides a deep implementation guide for sandboxing coding agents: enforce network egress and filesystem boundaries below the application layer, protect agent configuration files, isolate spawned hooks and MCP processes, use virtualization where warranted, inject scoped secrets, and expire sandbox state.
My takeaway: Start with three non-overridable controls: deny arbitrary egress, deny writes outside the workspace, and deny every agent write to executable configuration. Then sandbox the IDE and all child processes, restrict reads, broker short-lived task credentials, require fresh approval for each default-deny exception, destroy accumulated state on a defined schedule, and test the controls through indirect execution paths.