Toward an Equation that Anticipates AI Risks