Google Researchers Say AI Agents Need Privacy Rules That Follow Context

Published 29 October 2026 | Publisher: NewsPro | AI-assisted reporting based on the sources below.

AI agents can read information, call tools and perform sequences of tasks, so deciding whether an action is appropriate may require more than checking whether an app has permission. In an October 5 report, Google Research authors Eugene Bagdasarian and Marco Gruteser argue that privacy and security controls should consider the context in which information moves and actions occur.

The Google Research summary says the report followed a workshop with more than 50 academic and industry participants. It draws on contextual integrity, a framework that evaluates information sharing in relation to the people involved, the information type and the rules governing its transfer. Helen Nissenbaum’s foundational paper explains why privacy cannot be reduced to secrecy: sharing may be appropriate in one setting and inappropriate in another.

Why static permissions can fall short

An agent may need calendar details to arrange a meeting, but that does not authorize sending them to every service it encounters. The report argues manually written rules are difficult to scale across long-running tasks and changing tools.

Its proposed research directions include a supervisor layer to check actions against contextual policies, sandboxing that limits agent access, models that reason about unclear instructions, and user interfaces that adapt as circumstances change. It also identifies challenges in agent-to-agent collaboration and in agreeing on norms across organizations.

An agenda with open questions

Open questions include how to encode social expectations without blocking useful work, how users correct policies and who decides which norms apply. Model judgment alone cannot guarantee correct behavior.

This is a research agenda, not a new Google product or evidence that contextual policy engines are already preventing privacy failures in deployed agents.