OpenAI explains how the Codex sandbox limits file and network access for local commands
Inside the Codex sandbox
The Codex sandbox runs commands inside the user's existing environment and uses OS mechanisms to restrict filesystem and network access according to permission settings. The built-in :workspace profile allows workspace writes and broad reads, protects directories like .git and .codex, and disables network by default, but it can still read sensitive files or modify local dependencies. OpenAI details the macOS, Linux and Windows implementations and shows how codex sandbox applies permission limits to homegrown tools and MCP services.
Why it matters: OpenAI shows how permission profiles and the sandbox command can constrain MCP services, reusing file-write and network limits for tools beyond the agent itself.