OpenAI running Codex safely — sandboxing, approvals, and telemetry

AI Impact Summary

OpenAI employs a multi-layered security approach for Codex, utilizing sandboxing to isolate model execution, stringent network policies to restrict access, and agent-native telemetry to monitor and audit coding agent activity. This strategy is designed to mitigate the risks associated with Codex's powerful code generation capabilities, ensuring compliance and preventing unauthorized usage. The combination of these controls allows OpenAI to support the broader adoption of coding agents while maintaining a high level of security.