New research proposes a framework for evaluating general-purpose models against novel threats