OpenAI has outlined how its Codex coding agent runs commands directly on a user's computer while keeping those actions within limits the user sets. The episode focuses on the sandbox that sits between the agent and the machine, which controls what files Codex can read or change and whether it can reach the network. The design addresses a central tension in AI coding tools. An agent that can run shell commands is useful because it can build, test and edit code without constant supervision, but the same access could expose files or send data to external servers if the agent makes a mistake or is manipulated. Filesystem and network restrictions are intended to narrow that risk, with the boundaries defined by the user rather than left to the model. The sandbox is described as working across macOS, Linux and Windows, which matters for developer teams that run mixed operating systems. For Australian businesses and professionals, this changes little on its own, but it is relevant if your developers already use Codex or similar agents. Any tool that can read a codebase may touch personal information held in test data, logs or configuration files, so teams should check that sandbox settings block access to those locations and that network access is limited where needed, with the Privacy Act 1988 in mind. It is worth trying in a non-production project first, with the most restrictive settings, before extending access. If your organisation does not use Codex or any AI coding agent, there is no action required.