OpenAI’s Codex agent can autonomously manage code repos and run commands, but with strict security guardrails. The system operates in a sandbox that limits what it can access, and an approval policy determines when human sign-off is needed for higher-risk actions. A new Auto-review mode lets the agent approve routine, low-risk tasks on its own—like running tests or checking dependencies—while still requiring user permission for anything outside its safety zone. Network access is managed, blocking unknown destinations, and credentials are stored securely in the OS keyring tied to ChatGPT enterprise controls. The goal is to let developers move fast on everyday tasks while maintaining clear boundaries for security teams. Why it matters: This framework shows how the industry is tackling the tension between agent autonomy and enterprise security—a template that will likely shape how other AI coding tools handle risk as they get deployed at scale.