Google Research Flags 3 Agentic AI Risks in Report Backed by 50-Plus Experts

3 articles · Updated · Google Research · Oct 5

Google Research published a workshop report arguing that autonomous AI agents need new privacy and security safeguards because they may access personal data and take consequential actions across many contexts.
The report says agentic systems differ from traditional software in three critical ways and cannot rely on static permissions alone, making context-sensitive judgments about appropriate data sharing and actions central to trust.
Google proposes a supervisor layer with a contextual policy engine that would generate real-time rules for each task, checking whether data flows or actions are appropriate before information leaves a user's workspace.
More than 50 academic and industry participants contributed to the report after a late-2025 New York workshop, which also calls for standardized multi-agent benchmarks and open-source "Agent Gym" test environments.
The broader aim is to turn contextual integrity—a privacy theory focused on appropriate information flows—into a foundation for safer, more trustworthy agent ecosystems across academia, industry and government.