OpenAI details Codex sandbox design using native OS isolation instead of containers
·2026.11.18 17:00
Key point
The system restricts file and network access via Seatbelt, bubblewrap, and MXC to prevent data exfiltration while preserving local toolchains.
Details
- Codex applies a defense-in-depth sandbox to agent commands, using native OS mechanisms rather than VMs or Docker to preserve local development environments like iOS toolchains.
- The default
:workspaceprofile allows broad file reads and writes within the workspace but blocks network access and keeps.gitand.codexdirectories read-only to prevent configuration tampering. - A shared
spawn_command_under_sandbox()API enforces permissions across macOS, Linux, and Windows, with child processes inheriting constraints from the parent process tree. - Network isolation is enforced via a proxy that filters requests against domain rules, pausing unmatched requests for user approval or auto-review to prevent data exfiltration.
Platform Implementation
- macOS uses Apple's Seatbelt (SBPL) via
sandbox-exec, starting with adeny defaultpolicy and allowing connections only to local proxy ports. - Linux combines
bubblewrapfor mount namespace isolation withseccomp-bpffor syscall filtering, using a host bridge to relay network traffic through a loopback listener. - Windows is transitioning from legacy ACL-based sandboxes to the Microsoft eXecution Container (MXC) library to eliminate the need for elevated privileges and persistent host state.
Configuration and Limitations
- Permissions are defined in
config.tomlprofiles, where explicitdenyrules overrideallowrules, and managed requirements can restrict available profiles. - The
codex sandboxCLI allows testing arbitrary commands under specific profiles without an active agent session. - stdio-based MCP servers are not sandboxed by default because the harness cannot reliably infer their resource needs, requiring manual wrapping with specific permission profiles.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.