OpenAI Details Codex Sandbox: OS-Level Isolation Over Containers

On November 19, an OpenAI developer account published a series of long posts detailing the Codex sandbox design: instead of moving computation into separate containers, Codex constrains command execution directly in the user's existing environment with OS-level sandboxes — Seatbelt on macOS, bubblewrap on Linux. The company also shared details of the default permission profile: network access is disabled, with commands requesting network triggering a user prompt or routing to auto-review; when a Git repository is the workspace, source files can be edited without approval, but writes to the root-level .git and .codex folders are blocked, broad filesystem reads are allowed, and all subprocesses inherit the sandbox.

Confirmed

Why it matters

2026-11-19 ~ 2026-11-19 · 5 related posts

Primary sources