The development

What happened

OpenAI has detailed how Codex, its AI coding harness, runs commands in a developer's existing environment while applying filesystem and network restrictions through operating-system controls. The 18 November engineering post describes permission profiles shared across macOS, Linux and Windows, and a transition to Microsoft's MXC isolation library on compatible Windows devices. It is an architectural account, not proof that every installation has the same protections or that sandboxing removes all risk.

Key facts

  • The described workspace profile allows project writes while protecting paths such as .git and .codex, permits broad reads subject to OS permissions, and disables networking. Broad reads can still include sensitive files; writable local dependencies can create risks outside a command's sandbox.

  • Named profiles define read, write, deny and network rules. Enterprise-managed requirements can constrain selectable profiles. A model can request additional access, but the harness applies the user's approval policy; requesting escalation does not grant it.

  • Domain restrictions depend on an enabled network proxy, not domain rules alone. OpenAI says the proxy decides whether a host is allowed before DNS resolution. Broad wildcard rules or access to DNS services can still enable data leakage.

  • The post describes Seatbelt on macOS and bubblewrap namespaces plus seccomp filtering on Linux. Child processes inherit restrictions. Linux and MXC deny globs are expanded before launch, so new matching paths are not automatically added while that command runs.

  • The Windows implementation is transitioning to MXC on compatible devices, not described as universally complete. Its configured controls disable clipboard access and input injection, but host-loopback services remain reachable: isolation boundaries are not identical across platforms.

  • Stdio MCP server startup commands are not automatically sandboxed by Codex, the post says. Its separate sandbox command can wrap a server or another tool with a chosen profile. The article's examples are explanations, not a TideReach-tested configuration.

The context

Why it matters

TideReach assessment: an AI agent's permission request, the harness's approval decision and operating-system enforcement are separate controls. The practical boundary also depends on what remains readable or writable and which tools run outside it. OpenAI's account makes those trade-offs clearer, but does not establish an independently audited security guarantee or justify relaxing permissions. TideReach has not installed, exercised or penetration-tested these configurations.