Anthropic's October 2024 Responsible Scaling Policy update tied stronger safeguards to specific increases in model capability. Its announced framework separated two questions: whether a model could perform a sufficiently dangerous task, and whether protections were adequate for that capability. Thresholds covered advanced autonomous AI research and assistance with chemical, biological, radiological or nuclear weapons. The proposed checks therefore addressed both the model and the defenses around it. This was a governance commitment describing when protections should increase, rather than a measured demonstration that those protections would prevent every misuse.
