DeepMind's safety framework is based on so-called "critical capability levels" (CCLs). These are essentially risk assessment rubrics that aim to measure an AI model's capabilities and define the point ...