In our view, developers of AI models — working with safety and security experts — need to specify which AI capabilities are most likely to lead to this kind of pandemic-scale harm.