Authorization Before Context: New AI Security Model Prevents Cross-Audience Memory Leakage in Agentic Systems

An arXiv paper introduces 'authorization before context,' a security rule for personal language agents to stop memory leakage between audiences. The system tags each stored item with the audience present at recording. When assembling a prompt, it reads the current viewer set from channel metadata, falling back to public if ambiguous. The item is admitted only if every current viewer already belonged to its original audience. This anti-monotone membership rule ensures participants retain cross-channel recall while excluding narrow-audience content from broader viewers, not relying on model behavior. The paper highlights memory-to-context transition as a vulnerable attack surface, including ambiguous channels, cross-audience prying, and poisoned memory.