Indirect prompt injection (IPI) poses a significant and evolving security threat to AI applications, particularly those like Workspace with Gemini. Attackers can inject malicious instructions into data used by the LLM, influencing its behavior without direct user input. Google addresses IPI through a multi-faceted and continuous defense strategy. This involves proactively discovering and categorizing new attack vectors using internal and external programs. Human and automated red-teaming exercises simulate attacks and test for vulnerabilities. The Google AI Vulnerability Rewards Program (VRP) allows collaboration with external security researchers. Open-source intelligence feeds help track publicly disclosed AI attacks. Newly discovered vulnerabilities undergo thorough analysis and are added to a vulnerability catalog. Synthetic data generation expands attack scenarios for comprehensive testing. Google employs deterministic defenses for rapid response and ML-based defenses via model retraining. LLM-based defenses are improved through prompt engineering. Model hardening enhances the Gemini model's resilience. Defense effectiveness is measured through simulated attacks and comparative testing. Google is committed to a secure and trustworthy AI experience by combining security research, automated pipelines, and advanced ML/LLM models. This iterative framework ensures they stay ahead of evolving threats in the IPI landscape.