OpenAI described its Dots safeguards on September 29. A separate system called Auto-review checks planned actions against instructions and safety requirements.
The provider says proactive background research uses tools technically restricted to reading. Active tasks are not universally read-only.
Some actions require approval every time; password changes and financial transfers require human handoff. OpenAI acknowledges remaining mistakes.
This risk signal documents announced safeguards, not independently proven reliability or a new incident.