Meta extended its loss-of-control rules to AI training and evaluation on October 2, 2026. The revised Superintelligence Scaling Framework addresses risks before a model is released.
For high-risk runs, Meta specifies internally vetted sandboxes and protected real-time logs. Automated monitoring is intended to detect severe misconduct and escape attempts and stop individual rollouts.
The framework nevertheless allows approved monitoring exceptions. Meta describes shutoff systems as a development goal; their effectiveness has not been independently established.
Meta also plans an AI committee of its Board of Directors over the coming months. It would review changes and adherence to company rules. This is planned corporate oversight, not a government body.