OpenAI Details Codex Sandbox File and Network Limits

OpenAI has explained how the Codex sandbox restricts file and network access for commands and the processes they launch.

Updated on Nov. 19, 2026 in Artificial Intelligence

Isometric editorial illustration of a disconnected network plug and loose cable beside a folder block inside a boundary frame, depicting software access limits.
OpenAI has detailed how its Codex sandbox limits which files and network resources commands and their launched processes can reach during a session. AI Illustration. Upload story photo >

OpenAI detailed how its Codex sandbox limits command access to files and networks. The built-in :workspace profile blocks network access when a Git repository is the workspace root.

Why it matters

The sandbox sets the boundaries for what Codex can read, write and reach during a session, which determines how safely developers can let the tool operate on their code. Users and enterprises can customize these permissions and integrate the sandbox into their own tools.

The :workspace profile permits writes within the workspace while keeping .git and .codex read-only when a Git repository is the workspace root. Broad filesystem reads are allowed, subject to existing operating system permissions, and the sandbox enforces the same boundaries on macOS, Linux and Windows.

The players

OpenAI

OpenAI is an artificial intelligence research and deployment company that develops the Codex coding agent and its sandbox system.

The details

Permission profiles define the boundaries for each Codex session, and users can choose settings that allow changes in selected directories while disabling network access. For code review, a profile can keep the code under review read-only while still allowing documentation lookups, and commands plus the processes they start inherit the same sandbox boundaries on each platform.

Under Permission Profiles

The Codex sandbox operates through named permission profiles, with the :workspace profile as its built-in default that blocks network access when a Git repository is the workspace root. This places the sandbox within a model where AI coding agents receive explicit, configurable file and network boundaries rather than unrestricted system access.

Developers using Codex can choose profiles that keep code read-only or block network access, which limits what the agent can change or reach on their machine. The same boundaries apply across macOS, Linux and Windows, so settings carry over regardless of platform.

The takeaway

Developers who run Codex should review the active permission profile before starting a session, especially when working in a Git repository with network access that they do not need. Matching the profile to the task keeps the agent's reach limited to what the work requires.

Further reading

Explore more Artificial Intelligence coverage for context on how AI coding tools are being constrained.

Source note: This article includes information reported by TokenPost.