Taxonomy, Evaluation and Exploitation of IPI-Centric LLM Agent Defense Frameworks