CAMP Framework Addresses Cumulative Privacy Risks in Multi-Turn LLM Conversations

A recent study presents CAMP (Cumulative Agentic Masking and Pruning), a framework aimed at tackling privacy risks in multi-turn dialogues involving Large Language Models. This research, available on arXiv with the identifier 2604.16521v1, reveals the inadequacies of current Personally Identifiable Information (PII) protection methods in agentic conversations. Existing PII masking techniques treat each message independently, replacing identified entities with placeholders before the sanitized text is sent to the model. While these methods effectively prevent direct identifier leaks within single messages, they fall short in managing the increased privacy threats posed by accumulating PII across multiple exchanges. The study defines this issue as Cumulative PII Exposure, illustrating that users can inadvertently create a complete re-identifiable profile by sharing their name, employer, location, and medical information across different messages. The introduction of LLMs in agentic, multi-turn dialogues has uncovered privacy vulnerabilities that current protection strategies do not adequately address. The CAMP framework marks a crucial step forward in enhancing privacy safeguards for conversational AI systems.