This paper models the stochastic behavior of large-scale terrorism using extreme value methods. We utilize a unique dataset composed of roughly 26,000 observations. These data provide a rich description of domestic and international terrorism between 1968 and 2006. Currently, a credible worst-case scenario would involve losses of about 5000 to 10,000 lives. Also, the return time for events of such magnitude is shortening every year. Today, the primary threat is from conventional weapons, rather than from chemical, biological and/or radionuclear weapons. However, pronounced tails in the distribution of these incidents suggest that this threat cannot be dismissed.