Google Ships Zero-Trust AI Agent Blueprint with ADK

Google open-sourced a zero-trust AI agent example built on ADK and Gemini for customer service/returns, defending against prompt injection. Key point: security lives outside the LLM context via hardware-backed cryptographic signing for irrefutable DB writes, gVisor sandboxing for code execution, and a deterministic semantic gateway for business logic checks — system prompts alone aren't treated as a security boundary.