← All AI news
OpenAISafety

Codex explains its sandbox—and a separate permission check for MCP servers

A new engineering explanation separates command restrictions from the permissions of connected tools.

THE UPDATE

What changed

OpenAI published an explanation of Codex permission profiles and their filesystem and network controls. It also notes that the harness does not automatically sandbox the command that starts a stdio MCP server.

One useful next step

Inventory one development workflow: writable folders, readable sensitive files, allowed network destinations and MCP launch commands. Ask the responsible administrator to verify each boundary in a disposable test environment before using private data.

Review permissions before connecting a tool ↗

Before you dive in

This is an engineering explanation, not evidence that your installation has these protections configured. Defaults, platform enforcement and managed policy matter. A sandbox reduces exposure but does not establish that generated code or tool behavior is safe; no configuration was tested for this brief.

GO TO THE ORIGINAL

Source & announcement date

OpenAI ·

OpenAI: Inside the Codex sandbox ↗ (opens in a new tab)

This brief summarizes the linked announcement. “The AI Superpower take” and the next step are our editorial interpretation, not claims from the source.

KEEP YOUR CURIOSITY GOING

A few more worth knowing.

All AI news ↗