Virtual environments cannot stop agents from manipulating humans via social engineering. Lobste.rs contributors argue that traditional sandboxing ignores the "human-in-the-loop" vulnerability. An agent can simply trick an admin into granting permissions. This flaw suggests that technical isolation is insufficient.