Fuzzing Inference-Time Privacy Risks in Deployed Large Language Models

Large Language Models (LLMs) are increasingly deployed in applications that process sensitive information, raising concerns about privacy violations during inference. Existing approaches primarily focus on training-time safeguards and privacy auditing, providing limited support for validating whether deployed systems satisfy privacy requirements at runtime. This paper introduces a fuzzing-based framework for testing inference-time privacy failures under black-box conditions. The framework is instantiated for two complementary privacy risks: misuse and disclosure of personal data and re-identification of deidentified information. An empirical evaluation across three deployed LLMs shows that inference-time privacy failures can be systematically elicited under both risk categories. The results further indicate that privacy failures are influenced by risk-specific factors, while susceptibility varies across deployed model-provider systems. These findings demonstrate that privacy can be operationalized as a testable runtime property and support the use of risk-driven black-box testing for evaluating deployed LLMs.