Efficiently Computing Susceptibility to Context in Language Models