Preventing Rogue Internal Deployments by AIs or Humans