“A potential safety measure for AI models with expert virology skills is to restrict access by default and require trusted users, such as researchers at Genentech, to request permission to have guardrails removed.”