Observable Channels, Not Just Storage: Evaluating Privacy Leakage in LLM Agent Pipelines

This paper redefines privacy leakage in LLM agent pipelines as a channel inversion problem. This novel formulation moves beyond specific system components, enabling a more generalized understanding of how hidden dependencies become observable signals. CIpl introduces a target independent framework to represent any LLM agent system through a common signature.