AI threats in the wild: The current state of prompt injections on the web

Google Threat Intelligence reports that Indirect Prompt Injection (IPI) is an active and prioritized threat against LLM-based agents. In a broad public-web sweep for known IPI patterns, they found real-world abuse where attackers inject malicious instructions into data or tools relied on by models—potentially influencing agent behavior or enabling data exfiltration even without direct user input. Google stresses IPI’s risk in agentic and multi-source AI applications (e.g., Workspace with Gemini), and describes a continuous, layered approach of monitoring, model hardening, and operational mitig