Determining LLMs’ Ability to Enforce Contextual Integrity in Privacy Decisions

Determine the extent to which large language models can reliably reason about and enforce Contextual Integrity norms—specifically the sender, subject, recipient, data type, and transmission principles—to make context-appropriate privacy decisions when operating as autonomous agents in open-ended environments.

Background

The paper notes that modern LLM agents access private data, process untrusted content, and communicate externally, creating powerful privacy leakage vectors. It emphasizes that current LLMs have limited capability to make context-appropriate privacy decisions and frames Contextual Integrity as the normative basis for judging appropriate data flows.

Given these challenges, the authors explicitly state uncertainty about how well LLMs can satisfy Contextual Integrity requirements and caution that, without establishing this capability, integrating agents broadly into users’ lives is premature. This motivates research to measure and validate LLMs’ abilities to apply CI parameters reliably in real-world scenarios.