Google Research proposes contextual policy engines and benchmarks for agentic privacy

Google Research published a report from its Contextual Agent Privacy and Security (CAPS) workshop, held in New York with over 50 academic and industry experts. It argues that agents need context-aware policies judging whether an information flow is appropriate, not just whether access is permitted, and calls for open Agent Gym environments to test multi-agent safety.

DEV Community

Open source

Report stems from the CAPS workshop with more than 50 academic and industry experts
Policies weigh actors, information type and transmission conditions, not just access rights
Open Agent Gym environments proposed to test how multiple agents behave together
No production product yet — it is an architectural direction, not a release

#Google

HANDBOOK.md benchmark: top AI agents follow policy rules only 36.2% of the time

Google Research proposes sequential auctions instead of a central orchestrator for multi-agent LLMs

Policy

Google Research proposes contextual policy engines and benchmarks for agentic privacy

HANDBOOK.md benchmark: top AI agents follow policy rules only 36.2% of the time

Google Research proposes sequential auctions instead of a central orchestrator for multi-agent LLMs

Google privacy policy lets Personal Intelligence data feed targeted ads

Researchers propose elastic transformers without retraining