Google Workspace’s continuous approach to mitigating indirect prompt injections

Google warns of an evolving threat called indirect prompt injection (IPI), where attackers influence LLM behavior by embedding malicious instructions into the data, tools, or content that a model ingests — potentially without any direct user input. IPI is highlighted as an ongoing, dynamic risk amplified by agentic automation and multi-source AI applications (e.g., Workspace with Gemini). Google says this is not a one-time fix and describes a continuous, comprehensive approach: improving model resistance to IPI, hardening AI application behaviors, and rolling out layered defenses and telemetry