Google DeepMind Researchers Warn Hackers Can Hijack AI Agents Through Malicious Web Content

Researchers at Google DeepMind present a systematic framework for “AI Agent Traps,” a class of adversarial web content designed to manipulate, deceive, or exploit autonomous AI agents. The paper categorizes six trap types—Content Injection, Semantic Manipulation, Cognitive State (RAG) Poisoning, Behavioural Control, Systemic, and Human-in-the-Loop—reports experimental success rates (in some cases exceeding 80%), demonstrates advanced techniques like dynamic cloaking and invisible prompt injection, and recommends layered defenses (model hardening, runtime filters, and ecosystem standards) while calling out unresolved legal accountability for agent-caused harms.