OpenAI explains Codex sandbox limits: workspace writes do not mean workspace-only reads

Read the original article →

OpenAI published an engineering explanation of the Codex sandbox on November 18, 2026.

What the default boundary does

OpenAI describes its built-in workspace profile as allowing workspace writes, protecting directories such as .git and .codex, permitting broad filesystem reads subject to operating-system permissions, and disabling network access. These are the documented profile’s defaults, not a description of every configured session.

Permissions can be tailored. A request for broader access is evaluated by the harness, not granted simply because the model asks.

A sandbox is not a correctness guarantee

OpenAI explicitly notes residual risk: writable dependencies could be modified and later executed by another tool outside the sandbox. Broad reads can also expose sensitive local files to a command even when network access is blocked. MintedBrain has not tested these protections.

MintedBrain analysis: match access to the task

Before an agent runs, identify the deliverable and the smallest set of files it needs to change. For a documentation edit, dependency installation and unrelated configuration changes deserve separate scrutiny. Afterward, inspect the diff and run checks appropriate to the actual change.

A useful review question is: “What could this command change now, and what might execute those changes later?” A successful build answers a different question from whether the access boundary was appropriate.

Continue with MintedBrain tutorials or Academy for learning workflows.

Source: OpenAI, Inside the Codex sandbox, November 18, 2026. Coverage checked November 22, 2026. Configuration and product behavior can change.

References

This article was originally published at OpenAI. For the full piece, read the original article.

Discussion

  • Loading…

← Back to News