It applies Contextual Integrity—the idea that privacy concerns appropriate information flows under social norms—to the question of whether an agent’s actions are appropriate in a particular setting.
The report identifies three characteristics that complicate conventional safeguards: agents interpret unstructured inputs, execute probabilistic plans, and act autonomously or delegate over longer tasks. It argues for coordinated research across system, model, user, and ecosystem levels. The post describes open problems and directions, not a deployed policy mechanism or evidence that contextual reasoning already makes agents safe.