The paper introduces AgentCIScope, a framework for evaluating the privacy behavior of LLM agents. According to the abstract, these agents are becoming more capable of handling daily workspace tasks by interacting with application environments through tool calls.
As agents read and write sensitive user data across email, documents, and other tools, the need for systematic evaluation grows. AgentCIScope grounds its assessment in contextual integrity, a theory that considers whether information flows are appropriate to the context.
The abstract does not provide details on the evaluation methodology or results, but the proposed framework appears aimed at addressing the privacy risks introduced by increasingly autonomous agents.