When Efficiency Backfires: Cascading LLMs Trigger Cascade Failure under Adversarial Attack

In three linesLLM cascade systems, designed to balance efficiency and performance by routing complex queries to powerful models, are vulnerable to targeted adversarial attacks. A novel attack exploits lightweight models and internal decision mechanisms to simultaneously degrade accuracy and cost-efficiency.

Summary generated by Claude — human-verified