Skip to main content
World Today News
  • Home
  • News
  • World
  • Sport
  • Entertainment
  • Business
  • Health
  • Technology
Menu
  • Home
  • News
  • World
  • Sport
  • Entertainment
  • Business
  • Health
  • Technology
OpenAI Details Codex Sandbox Security and OS-Based Isolation for Coding Agents

OpenAI Details Codex Sandbox Security and OS-Based Isolation for Coding Agents

November 18, 2026 Technology Desk Technology

Codex’s sandbox keeps network access off by default, a design choice OpenAI detailed in an engineering post published November 18 as the company explains how its coding agent stays confined on a user’s own machine.

The post, titled Inside the Codex sandbox and described in detail by BigGo Finance, lays out how the agent runs commands locally within filesystem and network limits the user controls. According to OpenAI, the sandbox is needed because running commands on the local machine is one of the most frequent — and one of the most potentially dangerous — operations Codex performs.

The mechanism matters now because agentic coding tools are spreading through enterprise engineering teams, and security teams are being asked to decide how much filesystem and network reach an autonomous agent should get.

“We take a defense in depth approach when it comes to safeguarding agents,” OpenAI stated in the post, noting that the sandbox restricts what commands are able to reach while training and alignment efforts are intended to ensure the model’s use of that access stays aligned with what the user wants.

Codex sandbox explained
Photo: ai-tracker.rootstudio.io
  • The Tech TL;DR:
  • Codex sandboxes commands using each operating system’s own isolation — Apple’s Seatbelt on macOS, bubblewrap and seccomp-bpf on Linux, Microsoft’s MXC on Windows — rather than a separate container.
  • The default :workspace profile allows writes inside the repository but keeps .git and .codex read-only, and blocks all network access until a user or profile allows it.
  • OpenAI acknowledged a real gap: an agent can plant code in an ignored directory like node_modules, where a formatter running on save could execute it outside the sandbox.

Codex uses host OS isolation facilities

The choice OpenAI made is one of architecture: instead of shifting execution into a dedicated container or virtual machine, Codex leans on the isolation mechanisms built into each operating system, letting the agent operate within the environment the user has already set up. The practical payoff, as BigGo Finance reported, is that commands can still use host tools such as an iOS build toolchain and simulator on a Mac.

The sandbox covers spawned commands, not only Codex’s own file operations, according to the sandbox documentation cited by Sulat AI & Tech News — so git, package managers and test runners inherit the same limits. Each platform enforces the boundary its own way.

Openai codex sandbox os boundary
Photo: news.sulat.com
Platform Isolation mechanism Notable behavior
macOS Apple’s Seatbelt, invoked via /usr/bin/sandbox-exec Policies written in Sandbox Profile Language; base profile starts by denying everything
Linux bubblewrap plus seccomp-bpf Vendored copy of bwrap; blocked network syscalls when access is disabled
Windows Microsoft’s MXC library (in the middle of a transition) Process isolation handled natively; standing up the sandbox requires no elevated access

On macOS, OpenAI said that Apple leaves the Sandbox Profile Language undocumented for third-party developers and that sandbox-exec carries a deprecation marker, yet macOS continues to bundle Seatbelt profiles and applications such as Google Chrome and Chromium still depend on it. Codex plans to keep using Seatbelt for the foreseeable future.

On Linux, the harness sets the no_new_privs flag so child processes cannot gain privileges. If the platform sandbox cannot enforce the selected policy on macOS, Codex refuses to run the command rather than running it unsandboxed, per the permissions docs Sulat cited.

The Windows transition is the newest piece. OpenAI wrote up that side on May 13 in a post on building a safe, effective sandbox for Codex on Windows, Sulat reported, saying Windows does not provide this kind of isolation out of the box. The old setup required administrator access, ran slow permission checks and left persistent state on the host — problems the MXC-based replacement is meant to remove.

Default profile allows broad filesystem reads

When a Git repository is the workspace root, the built-in :workspace profile permits writes inside the workspace but keeps .git and .codex read-only.

It allows broad filesystem reads subject to the user’s existing OS permissions, because compilers, runtimes and build tools depend on libraries and configuration outside the repository. That also means files such as API-key configs elsewhere on the machine may stay readable, as AI Tracker’s summary of the post noted.

As models become more capable, limits on what they can access matter more. A command editing your repo shouldn’t automatically be allowed network access. The Codex sandbox lets Codex run commands in your existing environment within the filesystem and network limits you control. pic.twitter.com/2qyoFYggPH

— Post by @OpenAIDevs on X

Codex ships three built-in profiles — :read-only, :workspace and :danger-full-access — and users can define their own under [permissions] in config.toml, extending a built-in so protected paths stay read-only unless overridden. For networking, the options are disabled entirely, unrestricted, or channeled through a proxy governed by allow and deny domain rules, with explicitly denied domains winning over allows.

Network enforcement is the sharpest edge of the design. A proxy checks each request against domain rules, and requests to unlisted domains can be paused for user approval or routed to automated review.

Only once the proxy has determined a hostname is allowed does it resolve it, since even a DNS lookup could expose information hidden in the name itself. OpenAI cautioned that broad wildcard rules or access to DNS services can still permit exfiltration.

# Example: a custom profile extending :workspace in config.toml
[permissions.my-profile]
extends = ":workspace"
network.enabled = true
network.allowed_domains = ["registry.npmjs.org"]
filesystem.deny = ["~/.aws", "~/.ssh"]

The node_modules problem OpenAI itself flagged

OpenAI did not present the sandbox as airtight. As one example, an agent might hide harmful code inside a locally installed package like node_modules/prettier; since Git typically overlooks that directory, the modification could go unnoticed in git status, and an editor configured to run the formatter on save might run that code beyond the sandbox’s reach.

Administrators get their own layer. Managed requirements files and cloud-supplied policy can constrain which profiles users may select, with cloud rules overriding local ones — a structure that lets a security team enforce a floor on agent permissions while individual developers adjust settings within it.

A race with the platform vendors

The post arrives amid parallel moves. With Apple’s macOS 27.3 Beta 2, the company introduced alerts that appear when applications seek full disk access, along with more granular choices for approving folder-level permissions — a step that came after a developer notice highlighted the dangers posed by ever more autonomous AI agents, BigGo Finance reported.

Microsoft has proposed an Agent Workspace in Windows 11 that gives agents a separate account and a restricted working area.

OpenAI also made the mechanism available for others to use: through a codex sandbox command, developers can take the identical configuration and platform-level enforcement into their own tools, and the company demonstrated wrapping an MCP server in a profile that shuts off network access and restricts writes to a cache and data directory.

OpenAI Developers put the framing plainly in a thread on X on November 19: “As models become more capable, limits on what they can access matter more. A command editing your repo shouldn’t automatically be allowed network access.” The company said that it aims to expand what agents are able to accomplish while ensuring people retain authority over what those agents can access.

Share this:

  • Share on Facebook (Opens in new window) Facebook
  • Share on X (Opens in new window) X

Related reading

  • Nvidia Interim CEO Colette Kress Promises December Pay Raises, Equity at First All-Hands Since Jensen Huang’s Resignation
  • How to Lower Your Internet Bill by Calling the Retention Department
  • 3.6% for Social Security, 2.6% for FERS, and a Bill Still Pending for VA (daybreakwire.com)

Related

Search:

World Today News

World Today News is your trusted source for global journalism — breaking headlines, in-depth analysis, and reporting from around the world.

Quick Links

  • Privacy Policy
  • About Us
  • Accessibility statement
  • California Privacy Notice (CCPA/CPRA)
  • Contact
  • Cookie Policy
  • Disclaimer
  • DMCA Policy
  • Do not sell my info
  • EDITORIAL TEAM
  • Terms & Conditions

Browse by Location

  • GB
  • NZ
  • US

Connect With Us

© 2026 World Today News. All rights reserved. Your trusted global news source directory.
For contact, advertising, copyright, issues email: [email protected]

Privacy Policy Terms of Service