Google DeepMind Publishes First Complete Taxonomy of AI Agent Attacks: Six 'Trap' Categories Map How Malicious Web Content Hijacks Autonomous Agents

SecurityWeek·high signal

Google DeepMind researchers published the first comprehensive taxonomy of web-based attacks against autonomous AI agents on April 1, 2026, gaining significant coverage this week. The six categories: Content Injection Traps (hidden instructions in HTML comments/CSS), Semantic Manipulation Traps (biased framing that skews reasoning), Cognitive State/Memory Traps (RAG knowledge poisoning), Data Exfiltration Traps (80%+ success rate across five tested agents), Systemic/Multi-Agent Traps (synchronized market crashes), and Human-in-the-Loop Traps (approval fatigue attacks). The paper identifies a critical Accountability Gap — when a compromised agent commits a financial crime, legal liability between operator, model provider, and domain owner is unresolved.