Google Deepmind's "AI Agent Traps" paper maps 6 attack types targeting autonomous AI agents, with exploit rates reaching 86% in tests.
