OpenAI makes Auto-review free in Codex, the second agent that checks risky actions
IA en un minuto newsroom · Editor: Jon Elgezabal

In 30 seconds
Auto-review, the automatic checker in Codex, no longer costs anything: anyone with a ChatGPT account can use it without spending their plan's usage. When the agent wants to leave its sandbox, for instance to go online, a second agent decides whether it may proceed and why. According to Thibault Sottiaux, it only aims to stop risky moves and anything that strays from the user's request.
OpenAI has made Auto-review free. It is the Codex feature (Codex is the OpenAI coding agent) in which a second agent reviews what the first one does. It is now free for everyone who signs in with a ChatGPT account and no longer draws usage from their plan. Thibault Sottiaux, who works on Codex and ChatGPT, announced it in a post on X on Tuesday, October 6.
Codex works inside a set of limits, the sandbox (the bounded space where it can act without asking for permission). When it needs to go beyond them, for example to reach the network or edit files outside the allowed folders, it asks for approval. With Auto-review, that request does not wait for a person: another agent reviews it, decides whether the action runs and explains why.
According to Sottiaux, its only goal is to prevent high-risk actions and actions that do not match what the user originally asked for. That makes it possible to run long tasks without approving everything by hand, which he says leads to decision fatigue unless a lot of time is spent configuring specific rules. It can be turned on in Settings > Permissions > Auto-review.
OpenAI states in its documentation that Auto-review does not give the agent more permissions, it only changes who reviews actions that already needed approval, and that it is not a deterministic security guarantee. Its help center confirms that, when users are signed in with a ChatGPT account, Auto-review safety checks are free and do not count toward the plan's usage limits.
Why it matters · analysis and opinion
Free and no longer drawing on the plan, Auto-review stops being an extra and becomes the sensible option for anyone running long tasks in Codex. It solves a concrete problem: approving every step outside the sandbox by hand is so tiring that it ends up being done without looking, and a separate reviewer that decides on each action and gives its reason is more useful than a distracted click. If it stops an action, Codex may look for a safer route or ask you. What matters is what it does not do. It does not give the agent more freedom, it only changes who reviews what already needed permission, and OpenAI warns that it does not guarantee absolute security. The sensible move is to start in a test repository until it is clear what it lets through and what it does not.
Source: Thibault Sottiaux (OpenAI) on X · Written with the help of AI: how we make the news


