CIPL: A Channel-Aware Framework for Recoverable Privacy Leakage in LLM Agents

TL;DR - CIPL is a channel-aware framework for measuring sensitive information that attackers can recover from black-box LLM agents, rather than merely detecting internal exposure. It enables consistent comparisons across memory-, retrieval-, tool-, and live-agent pipelines.

• Models leakage through source, selection, assembly, execution, observation, and extraction stages under a shared protocol.
• Memory leakage was nearly saturated, while retrieval-mediated leakage was often partial.
• Tool-mediated and live-agent leakage varied with observation surface, prompt-channel alignment, retrieval depth, and provider behavior.
• Semantic auditing identified useful disclosures missed by canonical exact-match metrics.