OpenAI has proposed adopting structured safety cases for frontier AI training, requiring rigorous containment, automated monitoring, independent executive veto powers, and public postmortems to prevent catastrophic model misalignment.
FireBrief
FireBrief