Auto-review and approvals
OpenAI says Auto-review checks actions that could affect accounts or share information, then allows, blocks, or asks again.
This guide restates “Action review and approvals” on Introducing dots and “Clear rules for taking action” plus “A separate check before dots act” on How we build safety, security, and privacy into dots. It does not add approval menus.
What Auto-review checks
On Introducing dots and How we build safety, security, and privacy into dots, OpenAI says Auto-review checks actions that could affect accounts or share information.
On those same pages, Introducing dots and the safety post, OpenAI says the check is against instructions, Custom Rules, and safety requirements.
Sharing and named authorization
On How we build safety, security, and privacy into dots, OpenAI says that to send a message or share a file, dots are taught to seek authorization that covers the information and the type of recipient.
On that safety post, OpenAI says health data always needs a named recipient. OpenAI’s example on that page is “share my medical history with Dr. Thompson.”
On that safety post, OpenAI says less sensitive personal data such as an email address or phone number by default needs a class of recipients. OpenAI’s example on that page is “any airline company.”
On that safety post, OpenAI says a Custom Rule can broaden less sensitive data. OpenAI’s example on that page is “share with any online form.”
On that safety post, OpenAI says authorization stays tied to the task; continuing later or delegating does not expand it.
If Auto-review allows a step
On How we build safety, security, and privacy into dots, OpenAI says that if Auto-review allows a step, the dot carries it out and continues.
On that safety post, OpenAI says it can rely on approval already given when that approval covers the action and the rules do not require a new confirmation.
If Auto-review blocks a step
On How we build safety, security, and privacy into dots, OpenAI says that if Auto-review blocks a step, the action does not run and the dot is told why.
On that safety post, OpenAI says the dot can ask for more information or approval, try a permitted alternative, hand a sensitive step back, or stop.
On that safety post, OpenAI says your approval cannot override core safety requirements.
Where the check sits
On How we build safety, security, and privacy into dots, OpenAI says the controls that enforce Auto-review sit outside environments the dot can change.
On that safety post, OpenAI says ordinary read-only steps still follow app permissions and other safeguards, but do not need this extra review.
Review consequential work
On Introducing dots and How we build safety, security, and privacy into dots, OpenAI says dots can still make mistakes.
On those pages, Introducing dots and the safety post, OpenAI says you should always review consequential work.
The safety post also points at Alignment for dots. OpenAI points here for safeguards and evaluations. This page does not summarize evaluation scores or section results.
What this page does not add
This page does not invent extra approval screens. For Custom Rules and confirmations you still make yourself, see Staying in control of a dot.