From Theory to Threat: Hackers Are Now Weaponizing Web Pages to Brainwash AI Agents

Unit 42 reports the first large-scale, in-the-wild occurrences of Indirect Prompt Injection (IDPI), a technique where attackers embed hidden instructions in web content so browser-integrated LLM agents and automated content pipelines execute attacker-provided prompts. Researchers observed IDPI used to bypass automated advertising safety checks and warn that as AI agents are integrated widely, attackers can scale impact by manipulating the data those agents consume; defenders must move beyond pattern matching to intent analysis and prompt-visibility telemetry.