The privacy risks associated with Large Language Models (LLMs) are increasingly critical. This talk examines two key areas: the assessment of privacy threats through membership inference attacks (MIAs), including current benchmarking challenges and bias issues; and the use of LLMs as privacy-enhancing tools, such as for anonymization or text rewriting to reduce attribute leakage. These dual perspectives highlight how LLMs can both compromise and protect privacy in AI systems.