Towards evaluations-based safety cases for AI scheming

Open in new window