OpenAI’s September 29 safety article describes a separate Auto-review check for actions such as sending email or changing files. The system compares planned actions with the user’s instructions, Custom Rules and safety requirements. A blocked step may lead to a clarification request, a permitted alternative or a handoff, rather than the action going ahead anyway.
The article says Custom Rules cannot remove mandatory safeguards. Certain sensitive steps, including password changes, remain with the user. It also explains that supported secure sign-in flows keep passwords outside the model’s context, while warning that a secret pasted into readable content may still be visible. OpenAI explicitly acknowledges that dots can make mistakes.
Editorial take: An assistant that occasionally pauses is not necessarily failing to be useful. The question is whether the pause clearly identifies the decision the person needs to make. Good delegation depends on both efficient execution and understandable boundaries, particularly when a task moves from preparing a draft to affecting another person or account.