Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle

Google Research has published a workshop report titled "Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle," described as a collaborative effort involving more than 50 academic and industry leaders from numerous institutions. The participants met at the Google Contextual Agent Privacy and Security (CAPS) Workshop, held in late 2025 in New York City. The report's authors include Eugene Bagdasarian, a research scientist, and Marco Gruteser, a principal scientist, both at Google Research.

The report identifies three dimensions in which AI agents differ from traditional deterministic software: unstructured interfaces and input ambiguity, which make expected behavior hard to define and complicate defenses against adversarial manipulation such as prompt injection; probabilistic control flows arising from generative planning, which traditional testing methodologies cannot easily secure; and autonomy and delegation, where longer tasks and sub-task delegation to other agents reduce the effectiveness of traditional user oversight and risk "confirmation fatigue."

The report is grounded in the theory of Contextual Integrity, which defines privacy as appropriate information flow according to social norms. It extends contextual integrity for information sharing to contextual security, described as the appropriateness of agent actions.

The report advocates a contextual policy engine within a supervisor layer to monitor and enforce action appropriateness, including a dynamic policy generation loop intended to operate in real time. It outlines opportunities across system-level sandboxing, model-level reasoning, user-centric controls, multi-agent interactions, and ecosystem governance. It also proposes standardized multi-agent benchmarks, described as dynamic "Agent Gym" environments, and calls for collaboration across academia, government, civil society, and industry. The post names Lillian Tsai as co-primary author and Sarah de Haas as coordinating workshop organization, report writing, and the blogpost.