A revised arXiv paper, PiSAs: Benchmarking Contextual Integrity in Multi-User Agentic Systems (arXiv:2607.05318v2), argues that LLM agents are moving from single-user assistants to shared organizational infrastructure. That shift, the authors say, creates new privacy risks.
The paper introduces PiSAs as a benchmark for evaluating these risks. Its central lens is contextual integrity: whether information flows appropriately depends on the context, not only on whether the information is sensitive. The abstract warns that inappropriate information may be exposed not just through outputs for external recipients, but also through other channels in shared agent workflows.
Because this digest draws on a single source, there are no independent findings to compare; the claims above reflect the paper's own framing. The benchmark appears aimed at researchers and builders of multi-user agent systems who need a more nuanced way to assess privacy failures.