The same AI capabilities underlie benefits and harms. How can we assure AI safety without curtailing beneficial applications? Defensive acceleration charts a possible course.