Explanation

Sensitive information disclosure in LLM systems can occur through prompt leakage, retrieval exposure, weak output filtering, or insecure context handling. Attackers coax models into revealing secrets, internal policies, or regulated records. Leakage risk rises when broad context windows and mixed-trust data are used without strict isolation.