Executive Summary

The article discusses two papers on contextual integrity in LLMs, focusing on the CIMemories benchmark. CIMemories uses synthetic user profiles with 100+ attributes and varied task contexts to test whether LLMs control memory‑based information flow appropriately. Evaluation shows frontier models leak up to 69% of attributes, with violations rising from 0.1% to 9.6% as tasks increase and to 25.1% over repeated prompts. Privacy‑conscious prompting fails; models over‑generalize, revealing a need for true context‑aware reasoning.

---

Intelligence Metadata - Source Publisher: Schneier on Security - Published Date: 2026-08-18T10:40:16+00:00 - Category: research

"Fortune befriends the bold."

— John Dryden