A Google Research workshop report published on Oct. 5 proposes a policy engine that would check whether an AI agent’s planned data sharing or action fits the task and its context. The document lays out open research problems; it does not report a tested system.
Agents can read personal data, call outside services and delegate work. A standing permission may grant access without spelling out where information should go in every new situation, while asking users to approve each step can become burdensome. The report treats that tension as a design problem for longer-running, more autonomous systems.
Its framework builds on “contextual integrity,” a view of privacy that asks whether information flows are appropriate for a particular setting, rather than assuming that all data must stay secret. The authors extend the idea to “contextual security”: whether an agent’s actions fit the expectations and norms of the situation.
The report sketches a two-part policy engine. One component would turn context and relevant norms into machine-checkable rules, asking the user to clarify intent when it cannot define a policy. A second would compare a proposed action with the agent’s current execution state, then allow, deny, modify or ask the user for confirmation. The report does not specify how that state should be represented or how such complex actions should be modeled. It uses conference-trip planning to show why appropriate information sharing may differ when booking flights, applying for a visa or contacting organizers.
That architecture remains illustrative. The report identifies unresolved problems such as translating ambiguous user intent into formal rules, resolving conflicting norms, resisting malicious instructions and making checks efficient at scale. It also calls for dynamic evaluations of extended, multi-agent tasks. It provides no experimental result showing that a contextual policy engine works or prevents privacy failures.
