This research introduces PrivaCI-Bench, a novel evaluation framework that assesses LLMs' privacy capabilities through contextual integrity and legal compliance lenses rather than traditional narrow definitions.

• Establishes a comprehensive benchmark that considers context, norms, and legal requirements when evaluating privacy
• Evaluates LLMs against established privacy regulations like GDPR and CCPA
• Provides a more nuanced understanding of how LLMs handle sensitive information across different contexts
• Offers practical insights for improving privacy protections in AI systems

This work matters for Security because it transforms how we evaluate and build privacy-aware AI systems, moving beyond simplistic approaches toward models that respect contextual norms and legal boundaries.

PrivaCI-Bench: Evaluating Privacy with Contextual Integrity and Legal Compliance