Running Codex safely at OpenAI

OpenAI has published a detailed breakdown of how it secures Codex, its AI-powered coding agent that can write and execute code autonomously. The company employs a layered safety approach: every code execution runs in a sandboxed environment isolated from the host system and network. This prevents unauthorized data access and limits damage from malicious or erroneous code. Additionally, Codex enforces granular approval workflows for sensitive operations—such as file system writes or network calls—ensuring human oversight before risky actions are taken. Network policies further restrict outbound connections to approved endpoints, reducing exfiltration risks.

To maintain transparency, OpenAI uses agent-native telemetry: every action Codex takes is logged and can be reviewed by administrators or security teams. This allows for real-time anomaly detection and post-incident forensics. The system also supports compliance requirements by recording the entire decision chain. By combining isolation, human-in-the-loop controls, and observability, OpenAI aims to enable enterprise adoption of AI coding agents without sacrificing security or regulatory compliance.