Results 131 - 140 of 23736

Frontier AI agents are increasingly delegated real work: browsing the web, calling APIs, and acting on user data. As autonomy grows, prompt injections or multi-modal perturbations can still lead agents to leak private data and execute harmful actions. I argue that we should rely on language models to decide what data and actions are appropriate in each context, but enforce these decisions using system primitives. Furthermore, decentralized multi-agent systems will allow us to increase robustness to attacks by providing independent perspectives on untrusted contexts. These contextual defenses offer a practical path to deploying trustworthy AI agents.