PersuasiveAgent: An Agentic Social Engineering Evaluation Framework to Assess Information Leakage of LLM-based Agents

Large language model-based agents are increasingly handles user’s sensitive data in applications such as banking, healthcare, and enterprise call centers. However, because they are built to be cooperative and prioritize helpfulness over defense, they end up vulnerable to psychological exploitation and manipulation that may release or expose sensitive information. We introduce PersuasiveAgent, an agentic evaluation framework that measures Personally Identifiable Information (PII) leakage under structured social engineering pressure. It uses Cialdini's six influence principles: Authority, Commitment, Liking, Reciprocity, Scarcity, and Social Proof, alongside a dynamic Bait-and-Switch strategy to assess if the underlying LLM exposes sensitive information.