Privacy Audit of Commercial Large Language Models with Sophisticated Prompt Engineering

Privacy audits of AI systems have become increasingly essential as they integrate more deeply into societal functions. The current investigation employs specialized prompt engineering to probe two commercial language models for potential privacy breaches and their compliance with the EU AI Act. Through a combination of qualitative and quantitative methods, including content, statistical, and cluster analyses, the study identifies significant privacy concerns related to data retention and leakage. Results indicate that revised prompting techniques uncover more pronounced privacy issues compared to standard methods, highlighting the need for continuous advancements in audit practices. This work not only maps the landscape of privacy vulnerabilities in contemporary AI models but also suggests actionable pathways for enhancing regulatory and development practices to safeguard user data.