Cascading Hallucination Attacks

An agent's tendency to generate plausible-but-false content is exploited so the fabrication propagates and amplifies through memory, self-reflection, or inter-agent communication rather than staying contained to one… In a multi-agent system: One agent misclassifies a financial-transaction anomaly as legitimate, and two downstream agents in the pipeline act on that classification without independently re-verifying it, propagating the wrong decision across the workflow.