OpenAI: Auto-review of agent actions without synchronous human oversight (alignment.openai.com)

OpenAI has announced the launch of Auto-review in Codex, a feature designed to enhance the autonomy of AI agents while maintaining safety standards without requiring synchronous human oversight. Unlike the previous Default mode, which frequently halts operations for human approval, Auto-review evaluates actions taken by Codex using a separate agent, significantly reducing the need for human intervention—stopping for approval approximately 200 times less and successfully approving around 99% of actions. This innovation not only streamlines workflows but also strengthens security by preventing unauthorized or damaging actions from being executed. The introduction of Auto-review holds significant implications for the AI/ML community as it strikes a balance between user autonomy and safety, addressing challenges associated with existing permission frameworks that hinder productivity. The system leverages GPT-5.4 Thinking to discern user intent and assess risks associated with external commands, ensuring a more reliable decision-making process. Although OpenAI acknowledges that Auto-review is not infallible and cannot provide absolute guarantees against malicious intent, it represents a critical step toward enabling AI systems to operate more independently while safeguarding against potential risks, thereby advancing the discussion around AI trust and safety.