OpenAI Details Controls, Boundaries, and Telemetry for Codex Coding Agents
Configure Codex with auto_review enabled and sandbox_workspace_write.writable_roots set to your development directories, and enable OpenTelemetry logging to capture agent activity.
Set up Codex with auto_review enabled and sandbox_workspace_write.writable_roots set to your development directories, and enable OpenTelemetry logging to capture agent activity.
Summary
OpenAI detailed the controls, boundaries, and telemetry used to govern Codex coding agents in real workflows on May 8, 2026. Codex operates inside a sandbox, with network policies that allow only cached web fetches and approved domains, and requires user approval for actions outside the sandbox.
The auto‑review feature automatically approves low‑risk requests, reducing friction for routine tasks while still stopping high‑risk actions. Rules define which shell commands are allowed or blocked, and CLI/MCP credentials are stored in the OS keyring with forced ChatGPT login tied to a specific workspace.
Codex emits OpenTelemetry logs for user prompts, tool approvals, execution results, and network decisions, which are available through the OpenAI Compliance Platform and can be exported to SIEM systems for audit and security triage.
Key changes
- Codex sandbox defines technical boundaries and requires approval for actions outside the sandbox
- Auto‑review automatically approves low‑risk requests to reduce user interruptions
- Network policy limits web fetches to OpenAI cache and approved domains, blocks unknown domains
- CLI and MCP credentials stored in OS keyring with forced ChatGPT login tied to a workspace
- Codex emits OpenTelemetry logs for prompts, approvals, execution results, and network decisions