There are many services which a site may wish to provide for its users, some of which may be external. There are a variety of security reasons to attempt to isolate services onto dedicated host computers. The services which a site may provide will, in most cases, have different levels of access needs and models of trust.
Services which are essential to the security or smooth operation of a site would be better off being placed on a dedicated machine with very limited access, rather than on a machine that provides a service (or services) which has traditionally been less secure, or requires greater accessibility by users who may accidentally suborn security.
It is also important to distinguish between hosts which operate within different models of trust (e.g., all the hosts inside of a firewall and any host on an exposed network). Some of the services which should be examined for potential separation are outlined in the section on service protection.
It is important to remember that security is only as strong as the weakest link in the chain. Several of the most publicized penetrations in recent years have been through the exploitation of vulnerabilities in electronic mail systems.
The intruders were not trying to steal electronic mail, but they used the vulnerability in that service to gain access to other systems. If possible, each service should be running on a different machine whose only duty is to provide a specific service. This helps to isolate intruders and limit potential harm.
There are two diametrically opposed underlying philosophies which can be adopted when defining a security plan. Both alternatives are legitimate models to adopt, and the choice between them will depend on the site and its needs for security.
The first option is to turn off all services and then selectively enable services on a case by case basis as they are needed. This can be done at the host or network level as appropriate. This model, which will here after be referred to as the "deny all" model, is generally more secure than the other model described in the next paragraph.
More work is required to successfully implement a "deny all" configuration as well as a better understanding of services. Allowing only known services provides for a better analysis of a particular service/protocol and the design of a security mechanism suited to the security level of the site.
The other model, which will here after be referred to as the "allow all" model, is much easier to implement, but is generally less secure than the "deny all" model. Simply turn on all services, usually the default at the host level, and allow all protocols to travel across network boundaries, usually the default at the router level.
As security holes become apparent, they are restricted or patched at either the host or network level. Each of these models can be applied to different portions of the site, depending on functionality requirements, administrative control, site policy, etc.
For example, the policy may be to use the "allow all" model when setting up workstations for general use, but adopt a "deny all" model when setting up information servers, like an email hub. Likewise, an "allow all" policy may be adopted for traffic between LAN's internal to the site, but a "deny all" policy can be adopted between the site and the Internet.
Be careful when mixing philosophies as in the examples above. Many sites adopt the theory of a hard "crunchy" shell and a soft "squishy" middle. They are willing to pay the cost of security for their external traffic and require strong security measures, but are unwilling or unable to provide similar protections internally.
This works fine as long as the outer defenses are never breached and the internal users can be trusted. Once the outer shell (firewall) is breached, subverting the internal network is trivial.