OpenAI shows how the Codex sandbox limits the access of an AI agent
Network access is off by default, but OpenAI tells that planted code can operate without the sandbox limits.
OpenAI published a report on the sandbox for Codex, its agent that writes code. A sandbox is an area with limits on the folders and network sites that software can get. It lets the agent operate commands on the computer of a user. OpenAI uses defense in depth, that is, more than one layer of protection.
How it works
Codex uses the safety tools that each computer system has. A permission profile sets the folder and network rules. In the workspace profile, the agent can write only in the project folder. A proxy checks each item of network traffic against domain rules.
- macOS: Seatbelt, which Apple marks as deprecated.
- Linux: bubblewrap and seccomp-bpf.
- Windows: a new sandbox on MXC replaces the sandbox that OpenAI uses at this time.
The proof
OpenAI gave no test results and no numbers. An agent can plant bad code in a package folder, which Git does not always show. An editor can then operate the code without the sandbox limits.
What is not known
OpenAI tells that large wildcard rules or access to DNS services can cause a leak of data. OpenAI does not tell the date of the new Windows sandbox.
Insight: Sandbox makers now compete in the open
In October 2026, Perplexity and AWS each made an open source sandbox for agents. A different company shows its sandbox and gives developers a command to use it on their tools. This reads as a move to set the standard wall for agents. The wall that developers use first can set the rules. Find which wall your agent tools use.
Sources
Posted