Type: url
Original: https://developers.googleblog.com/build-zero-trust-ai-agents-that-judge-intent-not-just-syntax/
Status: partial
Summary
Part 2 of a Google for Developers series on zero-trust agents, which keeps the Part 1 customer-support refund agent built with Agent Development Kit and moves its security checks from build-time code into managed runtime governance on Gemini Enterprise Agent Platform. It describes three controls applied through Agent Gateway — Model Armor screening prompts and responses at the edge, Semantic Governance Policies judging each proposed tool call against user intent and natural-language business rules, and Agent Anomaly Detection flagging multi-turn behaviour such as a refund drained across many turns — and a closed-loop remediation in which a detected anomaly leads to a new policy that takes effect without redeploying the agent.