“The primary security risk of AI agents comes from connecting them to tools, not from their intelligence in isolation.”