arXiv: Differential Privacy in Generative AI Agents — Formal Optimal Tradeoffs for Enterprise LLM Systems

arXiv·medium signalXBlueskyLinkedInCopy link

This paper formalizes the differential privacy tradeoffs when LLM agents access internal enterprise databases to generate context-aware responses, deriving provable optimal privacy-utility bounds for specific agent architectures and retrieval patterns. As enterprises deploy agents querying HR records, financial data, and customer PII, the gap between capability and formal privacy guarantees is becoming a compliance liability with no analytical framework to reason about it. The paper provides the first principled basis for engineers to specify what epsilon/delta guarantees are achievable given their deployment architecture.