LLMs and Contextual Integrity

Two new research papers examine how large language models handle contextual integrity—the ability to appropriately control what sensitive information gets shared depending on task context. Current frontier LLMs fail significantly at this challenge, with studies revealing up to 69% attribute-level violations where models leak information inappropriately, and violations that worsen and become unstable as usage increases. Proposed solutions using explicit reasoning prompts and reinforcement learning show promise in reducing information disclosure while maintaining task performance, though addressing contextual integrity appears to require fundamental improvements in how models reason about context rather than simple prompting techniques.