Privacy and Security of Frontier AI Agents

June 10, 2026 AI agents are increasingly tasked with complex tasks: operating on web APIs and user data. As autonomy grows, prompt injections and multi-modal perturbations could lead agents to expose that data or execute harmful actions. Leveraging the theory of Contextual Integrity, I argue that we should build language models with an understanding of social norms, use them to decide appropriateness in every context, but enforce these decisions using system primitives. Furthermore, while multi-party agentic systems could bring new threats, they could allow us to increase robustness to attacks