Running Codex safely at OpenAI

OpenAI has outlined its internal approach to running Codex securely, highlighting a deployment strategy built on sandboxing, approval policies, network controls, and agent-native telemetry. As AI coding agents increasingly take on autonomous tasks such as reviewing repositories and running commands, OpenAI emphasizes the need for technical boundaries that allow developers to move quickly on low-risk actions while stopping higher-risk tasks for explicit human review. To streamline this process, OpenAI utilizes Auto-review mode, where an auto-approval subagent evaluates routine requests and low-risk actions using recent context and user authorization levels. Network access is governed by managed policies that block unwanted domains and require approval for unfamiliar destinations, rather than granting open-ended outbound access. Additionally, authentication credentials are secured in the OS keyring, forced through ChatGPT, and tied to the enterprise workspace to integrate with compliance logging platforms. To address visibility challenges beyond traditional security logs, Codex supports OpenTelemetry log export for events like user prompts, tool approvals, and network decisions. OpenAI pairs these logs with an AI-powered security triage agent to analyze user and agent intent during unusual endpoint alerts. The same telemetry is also used operationally to monitor internal adoption and fine-tune rollout configurations across local surfaces such as the desktop app, CLI, and IDE extensions. Through these controls, OpenAI aims to provide security teams with the configuration management and agent-aware visibility needed for compliant and confident agent adoption.