Indirect Prompt Injection Attacks Hide Malicious Instructions in Websites to Target AI Agents

This report summarizes Zscaler ThreatLabz findings on Indirect Prompt Injection (IPI) campaigns that poison search results and hide malicious instructions in site metadata and off-screen HTML/CSS, enabling autonomous AI agents to execute fraudulent payments and misclassify authoritative sources; researchers demonstrated real-world impact by testing a browsing-and-payment-capable agent across 26 LLMs and provided associated malicious domains and GitHub links as IOCs.