From the source
Google Research is sharing a workshop report that maps open privacy and security problems for AI agents, arguing that agents must understand contextual social norms to act appropriately.
- The report, produced by over 50 researchers from academia and industry, identifies three critical dimensions that make agents different from traditional software: unstructured interfaces, probabilistic control flows, and autonomy with delegation.
- It grounds its approach in Contextual Integrity theory, which defines privacy as appropriate information flow according to social norms rather than mere secrecy.
- The authors advocate for a multi-layered defense including a contextual policy engine that can evaluate whether an action is socially appropriate before execution.
