arXiv cs.AIPaper
PrivDrift: Auditing User-Secret Leakage Under Topic Drift in Active LLM Conversations
This is a materialized risk for anyone building persistent assistants or shared-session systems. The finding is clear: topic drift does not reliably erase secrets, and leakage varies significantly by model. For builders: audit your deployment with this benchmark before your users disclose financial data or medical history in a shared session. For security teams: this is a harder problem than clearing context and needs architectural solutions, not prompt engineering.