We tested Context Bombs against seven AI models, measuring how deceptive documents, false leads and recovery tasks change attacker time, cost and completed objectives.