AI Agent Traps: Google DeepMind Maps Six Ways the Web Can Hijack Autonomous Agents

Google DeepMind published the first systematic taxonomy of adversarial attacks against autonomous AI agents. The paper identifies six trap categories — from hidden prompt injections (86% success rate) to memory poisoning (80%+ success at 0.1% contamination) — and calls them 'trivial to implement.' Here's what each trap does, why current defenses fail, and what the paper recommends.