AI TrackerGet the app
← All news
OpenAICoding

OpenAI explains how the Codex sandbox limits commands on macOS, Linux and Windows

In an engineering post, OpenAI's Michael Bolin describes how Codex runs commands on a user's machine within filesystem and network limits the user controls. Instead of moving work into a virtual machine or container, Codex uses each platform's own sandboxing facilities, so commands can still use host tools such as an iOS build toolchain and simulator.

  • Defaults: with a Git repository as the workspace, the built-in :workspace permission profile allows writes inside the workspace but keeps .git and .codex read-only, allows broad reads subject to the user's OS permissions (so files such as API-key configs elsewhere may stay readable), and disables network access. OpenAI says the sandbox is one layer of a defense-in-depth approach alongside training and alignment, and notes that harmful actions are still possible, for example tampering with an ignored node_modules package.
  • Permission profiles: users define named profiles under [permissions] in config.toml, can extend built-in ones, and set read, write or deny per path, including patterns relative to workspace roots. Network access can be off, fully open, or limited by a proxy with allow and deny domain rules, where deny wins.
  • Per-command requests: the model can ask for extra permissions or to run outside the sandbox, and the harness decides, by prompting the user or routing to auto-review.
  • Enterprise: administrators can constrain choices with managed requirements, for example allowed_permission_profiles; cloud requirements take precedence over /etc/codex/requirements.toml.
  • Reuse: a codex sandbox command lets other tools run commands, such as an MCP server, under the same sandbox.

The post also covers how the limits are enforced on macOS, Linux and Windows.

Get news like this as a push.

AI Tracker alerts you when something changes for the tools you use.

Get the app