Running Codex safely at OpenAI
A look at the controls, boundaries, and telemetry OpenAI uses to govern coding agents in real workflows.
TL;DR
- Codex agents are designed to act on behalf of users, performing tasks like reviewing repositories and running commands.
- OpenAI deploys Codex with controls such as managed configuration, constrained execution, network policies, and agent-native logs for safety.
- Sandboxing defines execution boundaries, while approval policies determine when user review is needed for actions outside these limits.
- Auto-review mode streamlines routine approvals for low-risk actions, reducing user interruption.
- Network access is managed, allowing only expected destinations and requiring approval for unfamiliar domains.
- Identity and credentials are secured through OS keyrings, forced login via ChatGPT, and workspace-level controls.
- Rules dictate command safety, allowing benign commands without approval while blocking or requiring review for dangerous ones.
- Agent-native telemetry, including OpenTelemetry logs, provides security teams with visibility into agent actions and intent.
- Codex logs are used with AI-powered security triage agents to distinguish expected behavior from malicious activity.