OpenAI has released a new sandboxing system that lets its Codex model run as a coding agent directly on Windows machines, according to a company blog post. The sandbox gives Codex controlled file system and network access, aiming to let it write, run, and test code locally without exposing the rest of the machine to risk.
How It Works
The sandbox creates an isolated environment for code execution on Windows, letting Codex build projects, run tests, and debug code while confined to defined boundaries. OpenAI says the system was designed to keep security overhead low enough not to hinder normal development workflows.
Key Security Features
According to OpenAI, the sandbox enforces several restrictions:
- Granular File Access: Codex can only access files within a specific, user-approved project directory, not personal or system files.
- Controlled Network Egress: Network requests are monitored and limited to prevent unauthorized communication or data exfiltration.
- Process Isolation: Processes spawned by the agent are contained within the sandbox and cannot interfere with other applications or system functions.
- Resource Limits: CPU, memory, and disk usage are capped to prevent the agent from consuming excessive system resources.
Why It Matters
Previously, running AI coding agents with real execution access was largely confined to cloud environments because of the security risks of unrestricted local access. OpenAI's sandbox is intended to make it safer for Codex to operate directly on a developer's Windows machine, extending agentic coding tools beyond cloud-only setups.