OpenAI Codex Brings a Windows-Native Sandbox for AI Coding Agents

Summary

In a sponsored demo, NetworkChuck shows Codex’s Windows-native agent sandbox using restricted identities, folder-scoped permissions, firewall-based network blocking and user approvals to limit what the AI agent can do.

Key points

• The video demonstrates Codex running natively in Windows, with a sandbox built on Windows security controls rather than a VM, Docker container or WSL2 environment.
• The sandbox uses a restricted identity with permissions scoped to a project folder; access to other directories can require user approval.
• By default, the demonstrated sandbox uses a no-network identity; Windows Firewall blocks an outbound request until the user explicitly approves it.
• The presenter cautions that full-access mode weakens these safeguards and says he does not recommend it for security.
• Codex runs PowerShell in the Windows environment, improving integration while making sandbox controls important for limiting agent actions.
• The demo also shows Codex creating Hyper-V virtual machines with Kali Linux and Ubuntu and an intentionally vulnerable app for a local penetration-testing lab.