TLDR Tech / InfoSec
Running Codex Safely at OpenAI
At OpenAI, Codex is deployed with the goals of keeping the agent inside clear technical boundaries, letting developers move quickly on low-risk actions, and making high-risk actions explicit. OpenAI deploys Codex in a tightly controlled sandbox and uses an auto-approval subagent for actions that cross the sandbox boundary, a tightly controlled network and authentication policy, and custom rules to allow fine-grained decisions on shell commands. Centralized configs and related are distributed to devices to standardize policies.
Read the full story