What Happened
On November 18, 2026, OpenAI described how Codex runs commands on a user’s machine and how users can control the filesystem and network limits applied to those commands across macOS, Linux, and Windows. [1]
OpenAI also announced a partnership to bring Le Figaro journalism to ChatGPT through summaries and excerpts linked to original stories. Separately, it reported using its models to optimize circuits in Jalapeño silicon and improve RISC‑V core designs. The chip-design account mentions GPT‑6 Astra, but does not identify a release date or an API change. Neither account provides a partnership or project date. [2] [3]
Why It Matters to Businesses
The Codex guidance is the most immediately actionable update for engineering teams: an agent that runs local commands needs explicit limits on what it can read, change, and reach over the network. The journalism partnership illustrates another product pattern—AI-generated summaries linked back to publisher content—but does not establish new access terms for other publishers. [1] [2]
Kimbodo Engineering Perspective
We would treat sandbox controls as one layer of an agent’s security design, not as a substitute for reviewing the commands, credentials, and data available to it. For the other announcements, we would separate demonstrated use cases from deployable capabilities: a chip-design result is not evidence that a new model version or API is available to customers. [1] [3]
How We Would Implement It
- Give coding agents a task-specific working directory, restrict filesystem access, and allow only the network destinations required for the job. Test those policies on each operating system in use. [1]
- Run agents with short-lived, least-privilege credentials; log commands and changes; and require human approval before merging or deploying agent-generated code.
- For publisher-backed summaries, retain links to original articles and review content-use permissions before adopting a similar workflow. [2]
Risks, Costs and Security
Tighter sandbox policies can interrupt legitimate builds that need dependencies or external services, so teams should budget for policy testing and exceptions. Looser policies increase exposure to unintended file changes, data access, and network activity. The announcements provide no pricing, API migration instructions, or dated GPT‑6 Astra release to plan against. [1] [2] [3]
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Application Development practice, or Estimate My AI Application.