In 1983 the Massachusetts Institute of Technology, working with IBM and Digital Equipment Corporation, embarked on an eight-year project designed to integrate computers into the university's undergraduate curriculum. The project was called Project Athena.
Athena began operation with nearly 50 time-sharing minicomputers: Digital Equipment Corporation's VAX 11/750 systems running Berkeley 4.2 Unix. Each VAX had a few terminals; when a student or faculty member wanted to use a computer, he sat down at one of its terminals.
Within a few years, Athena began moving away from the 750s. The project received hundreds of high-performance workstations with big screens, fast (for the time) processors, small disks, and Ethernet interfaces. The project's goal was to allow any user to sit down at any computer and enjoy full access to his files and to the network.
Of course there were problems. As soon as the workstations were deployed, the problems of network eavesdropping became painfully obvious; with the network accessible from all over campus, nothing prevented students (or outside intruders) from running network spy programs. It was nearly impossible to prevent the students from learning the superuser password of the workstations or simply rebooting them in single-user mode. To further complicate matters, many of the computers on the network were IBM PC/ATs running software that didn't have even rudimentary computer security. Something had to be done to protect student files in the networked environment to the same degree that they were protected in the time-sharing environment.
Athena's ultimate solution to this security problem was Kerberos, an authentication system that uses DES cryptography to protect sensitive information such as passwords on an open network. When the user logs into a workstation running Kerberos, that user is issued a ticket from the Kerberos server. The user's ticket can be decrypted only with the user's password; it contains information necessary to obtain additional tickets. From that point on, whenever the user wishes to access a network service, an appropriate ticket for that service must be presented. As all of the information in the Kerberos tickets is encrypted before it is sent over the network, the information is not susceptible to eavesdropping or misappropriation.
Windows 2000 includes a Kerberos Server, allowing Windows 2000 servers to act as authentication servers for Unix clients.
Kerberos authentication is based entirely on the knowledge of passwords that are stored on the Kerberos Server. Unlike Unix passwords, which are encrypted with a one-way algorithm that cannot be reversed, Kerberos passwords are stored on the server encrypted with a conventional encryption algorithm?in most cases, DES?so that they can be decrypted by the server when needed. A user proves her identity to the Kerberos Server by demonstrating knowledge of her key.
The fact that the Kerberos Server has access to the user's decrypted password is a result of the fact that Kerberos does not use public key cryptography. This is a serious disadvantage of the Kerberos system. It means that the Kerberos Server must be both physically secure and "computationally secure." The server must be physically secure to prevent an attacker from stealing the Kerberos Server and learning all of the users' passwords. The server must also be immune to login attacks: if an attacker could log onto the server and become root, that attacker could, once again, steal all of the passwords.
 Public key cryptography was not used because it was still under patent protection at the time that Kerberos was developed. There is a current IETF Internet-Draft entitled "Public Key Cryptography for Initial Authentication in Kerberos" that proposes methods for combining public key smart cards with Kerberos. This draft has been implemented in Microsoft's Kerberos.
Kerberos was designed so that the server can be stateless. (The server actually has a lot of permanent, sensitive state?the user passwords?but this is kept on the hard disk, rather than in RAM, and does not need to be updated during the course of Kerberos transactions.) The Kerberos Server simply answers requests from users and issues tickets (when appropriate). This design makes it relatively simple to create replicated, secondary servers that can handle authentication requests when the primary server is down or otherwise unavailable. Unfortunately, these secondary servers need complete copies of the entire Kerberos database, which means that they must also be physically and computationally secure.
Logging into a Unix workstation that is using Kerberos looks the same to a user as logging into a regular Unix computer. Sitting at the workstation, you see the traditional login: and password: prompts. You type your username and password, and if they are correct, you get logged in. Accessing files, electronic mail, printers, and other resources all work as expected.
What happens behind the scenes is far more complicated. When the workstation's login, sshd, or other network daemon (or, more commonly these days, the PAM library that performs authentication for these programs) knows about Kerberos, it uses the Kerberos system to authenticate the user. The details differ between the two different versions of Kerberos that are commonly available: Kerberos Version 4 and Kerberos Version 5.
 Patches for OpenSSH to use Kerberos 5 for authentication are available at http://www.sxw.org.uk/computing/patches/openssh.html. Although Kerberos 4 has also been used with SSH, it's much more difficult to make the two systems interoperate. Fortunately, the SSH protocol Version 2 can use the same security layer (GSSAPI) as Kerberos 5, which simplifies things considerably. The IETF Internet-Draft that covers the combination of these systems is draft-ietf-secsh-gsskeyex.
With Kerberos 4, the workstation sends a message to the Kerberos Authentication Server after you type your username. This message contains your username and indicates that you are trying to log in. The Kerberos Server checks its database and, if you are a valid user, sends back a ticket-granting ticket that is encrypted with your password. The workstation then asks you to type in your password and finally attempts to decrypt the encrypted ticket using the password that you've supplied. If the decryption is successful, the workstation then forgets your password and uses the ticket-granting ticket exclusively. If the decryption fails, the workstation knows that you supplied the wrong password and gives you a chance to try again.
 According to the Kerberos papers and documentation, there are two logical Kerberos Servers: the Authentication Server and the Ticket Granting Service. Some commentators think that this is disingenuous because all Kerberos systems employ a single physical server: the Kerberos Server, or Key Server.
 Actually, the initial ticket that the Kerberos Server sends your workstation is encrypted with a 56-bit number that is derived from your password using a one-way cryptographic function.
With Kerberos 5, the workstation waits until after you have typed your password before contacting the server. It then sends the Kerberos Authentication Server a message consisting of your username and the current time encrypted with your password. The Authentication Server looks up your username, determines your password, and attempts to decrypt the encrypted time. If the server can decrypt the current time (and the value is indeed current), it then creates a ticket-granting ticket, encrypts it with your password, and sends to you.
 Why the change in protocol between Kerberos 4 and Kerberos 5? Under Kerberos 4, the objective of the designer was to minimize the amount of time that the user's password was stored on the workstation. Unfortunately, this made Kerberos 4 susceptible to offline password-guessing attacks. An attacker could simply ask a Kerberos Authentication Server for a ticket-granting ticket for a particular user, then try to decrypt that ticket with every word in the dictionary. With Kerberos 5, the workstation must demonstrate to the Kerberos Authentication Server that the user knows the correct password. This is a more secure approach, although the user's encrypted ticket-granting ticket can still be intercepted as it is sent from the server to the workstation by an attacker and attacked with an exhaustive key search.
Figure 14-3 shows a schematic of the initial Kerberos authentication.
What is this ticket-granting ticket? It is a block of data that contains two pieces of information:
The session key
A ticket for the Kerberos Ticket Granting Service, encrypted with both the session key and the Ticket Granting Service's key
The user's workstation can now contact the Kerberos Ticket Granting Service to obtain tickets for any principal within the Kerberos realm?that is, the set of servers and users that are known to the Kerberos Server. Note that:
Passwords are stored on the Kerberos Server, not on the individual workstations.
The user's password is never transmitted on the network?encrypted or otherwise.
The Kerberos Authentication Server can authenticate the user's identity because the user knows the user's password.
The user can authenticate the Kerberos Server's identity because the Kerberos Authentication Server knows the user's password.
An eavesdropper who intercepts the ticket sent to you from the Kerberos Server will not benefit from the message because it is encrypted using a key (your password) that the eavesdropper doesn't know. Likewise, an eavesdropper who intercepts the ticket sent from the Kerberos Server to the Ticket Granting Service will not be able to make use of the ticket because it is encrypted with the Ticket Granting Service's key.
Once you have obtained a ticket-granting ticket, you will likely want to do something that requires the use of an authenticated service. For example, you will probably want to read the files in your home directory.
Under Sun Microsystems' regular version of NFS, once a file server exports its filesystem to a workstation, the server implicitly trusts whatever the workstation wants to do. If george is logged into the workstation, the server lets george access the files in his home directory. But if george becomes the superuser on his workstation, changes his UID to that of bill, and starts accessing bill's files, the vanilla NFS server has no mechanism to detect this trickery or to take evasive action.
The scenario is very different when NFS has been modified to use Kerberos.
When the user first tries to access his files from a Kerberos workstation, system software on the workstation contacts the Ticket Granting Service and asks for a ticket for the File Server Service. The Ticket Granting Service sends the user back a ticket for the File Server Service. This ticket contains another ticket, encrypted with the File Server Service's password, that the user's workstation can present to the File Server Service to request files. The contained ticket includes the user's authenticated name, the expiration time, and the Internet address of the user's workstation. The user's workstation then presents this ticket to the File Server Service. The File Server Service decrypts the ticket using its own password, then builds a mapping between the (UID, IP address) of the user's workstation and a UID on the file server. Figure 14-4 shows these operations.
As before, all of the requests and tickets exchanged between the workstation and the Ticket Granting Service are encrypted, protecting them from eavesdroppers.
The Ticket Granting Service was able to establish the user's identity when the user asked for a ticket for the File Server Service because:
The user's File Server Service Ticket request was encrypted using the session key.
The only way the user could have learned the session key was by decrypting the original Ticket Granting Ticket that the user received from the Kerberos Authentication Server.
To decrypt that original ticket, the user's workstation had to know the user's password. (Note again that this password was never transmitted over the network.)
The File Server Service was able to establish the user's identity because:
The ticket that it receives requesting service from the user is encrypted with the File Server Service's own key.
Inside that ticket is the IP address and username of the user.
The only way for that information to have gotten inside the ticket was for the Ticket Granting Service to have put it there.
Therefore, the Ticket Granting Service is sure of the user's identity.
And that's good enough for the File Server Service.
After authentication takes place, the workstation uses the network service as usual. Other Kerberized network services operate in a similar manner.
Kerberos is a general-purpose system for sharing secret keys between principals on the network. Normally, Kerberos is used solely for authentication. However, the ability to exchange keys can also be used to ensure data integrity and secrecy.
If eavesdropping is an ongoing concern, all information transmitted between the workstation and the service can be encrypted using a key that is exchanged between the two principals. Unfortunately, encryption carries a performance penalty. At MIT's Project Athena, encryption was used for transmitting highly sensitive information such as passwords but was not used for most data transfer, such as files and electronic mail.
In early 1996, graduate students with the COAST Laboratory at Purdue University discovered a long-standing weakness in the key generation for Kerberos 4. The weakness allows an attacker to guess session keys in a matter of seconds. A patch has been widely distributed; be sure to install it if you are using Kerberos 4.
 Incorporated into the CERIAS research center in 1998.
Kerberos has gone through five major revisions during its history. Both Kerberos 4 and Kerberos 5 are now in use.
Kerberos 4 is more efficient than Kerberos 5, but more limited. For example, Kerberos 4 can work only over TCP/IP networks. Kerberos 4 has not been updated in many years and is currently deprecated. In fact, some Kerberos 4 implementations are vulnerable to buffer-overflow attacks, and no patches have been posted.
Kerberos 5 fixes minor problems with the Kerberos protocol, making it more resistant to determined attacks over the network. Kerberos 5 is also more flexible: it can work with different kinds of networks. It also has provisions for working with encryption schemes other than DES. Although algorithms such as Triple-DES have been implemented, their use is not widespread largely because of legacy applications that expect DES encryption.
Finally, Kerberos 5 supports delegation of authentication, ticket expirations longer than 21 hours, renewable tickets, tickets that will work sometime in the future, and many more options. If you are going to use Kerberos, you should definitely use Version 5. IETF is working to revise and clarify RFC 1510, which defines Kerberos 5, and major protocol extensions are expected to follow.
Installing Kerberos is a complicated process that depends on the version of Kerberos you have, the kind of computer, and the version of your computer's operating system. It's a difficult task that requires you to have the source code for your computer system or the source code for replacement programs. It is not a task to be undertaken lightly.
Fortunately, you don't have to go through this arduous process. Kerberos or Kerberos-like security systems are now available from several companies, and they are a standard part of several operating systems, including Solaris, Mac OS X, and many Linux and BSD distributions. These days, there is no reason to be running anything but secure network services.
As an additional benefit, a version of Kerberos 5 has been included in Microsoft Windows from the Windows 2000 release onwards. Thus, with some effort, it is possible to make Kerberos interoperable between all your various Unix machines and Windows platforms.
 Note, however, that Microsoft has made proprietary modifications to the Kerberos protocol that force Windows clients to use Kerberos servers running on Windows servers. Microsoft's support of Kerberos merely allows Unix Kerberos clients to interoperate with Windows servers as well.
The MIT Kerberos source code is available from http://web.mit.edu/kerberos/www/. This is also the site where you can find official updates, patches, and bug announcements. There is also a free software implementation of Kerberos called Heimdal that is under active development; it is largely compatible with MIT's Kerberos, so most of what we discuss below should apply to Heimdal as well. You can get Heimdal at http://www.pdc.kth.se/heimdal/.
 Despite a weakening of U.S. export restrictions, MIT was, at the time of this writing, only willing to allow download by U.S. and Canadian citizens on advice from its legal counsel. An alternative source for MIT's Kerberos is http://www.crypto-publish.org/.
As the changes required to your system's software are substantial if you need to do them yourself, the actual installation process will not be described here. See the documentation provided with Kerberos for details.
Using a workstation equipped with Kerberos is only slightly different from using an ordinary workstation. In most implementations, all of the special Kerberos housekeeping functions are performed automatically. When the user logs in, the password typed is used to acquire a Kerberos ticket, which in turn grants access to the services on the network. Additional tickets are automatically requested as they are needed. Tickets for services are automatically cached. All of a user's tickets are automatically destroyed when the user logs out.
Of course, the Kerberos client needs to know where to find Kerberos servers. This can be configured manually on each client (traditionally in the krb5.conf file), or Kerberos servers can be advertised through DNS SRV records. IETF Internet Draft draft-ietf-krb-wg-krb-dns-locate describes this approach.
Kerberos isn't entirely transparent. If you are logged into a Kerberos workstation for more than eight hours, something odd happens: network services suddenly stop working properly. The reason for this is that tickets issued by Kerberos expire after eight hours, a technique designed to prevent a replay attack. (In such an attack, somebody capturing one of your tickets sits down at your workstation after you leave, using the captured ticket to gain access to your files.) Thus, after eight hours, you must run the kinit program, and provide your username and password a second time, to be issued a new ticket for the Kerberos Ticket Granting Service.
 A different window may be chosen at some sites.
Kerberos and LDAP
Kerberos mixes well with LDAP (discussed in the next section). Kerberos can be used to authenticate and secure LDAP queries and updates. Conversely, the LDAP database can store information about users that is more extensive than the data maintained by Kerberos alone, such as the user's home directory, shell, phone number, or other organizational information. Together, the two services can provide all of the functionality of NIS or NIS+, and they are increasingly being used to do so. Jason Heiss provides a good guide to this process on his page "Replacing NIS with Kerberos and LDAP" at http://www.ofb.net/~jheiss/krbldap/.
LDAP is sometimes used to store Kerberos keys. The Windows implementation of Kerberos uses Microsoft's Active Directory Service (a flavor of LDAP) to store Kerberos keys. Heimdal Kerberos supports this functionality. MIT Kerberos does not, out of concern that sensitive security infrastructure should be centralized at the Kerberos Server, rather than distributed via LDAP.
Although Kerberos is an excellent solution to a difficult problem, it has several shortcomings:
Because of the Kerberos design, every program that uses Kerberos must be modified. The process of performing these modifications is often called "Kerberizing" the application. The amount of work that this entails depends entirely on the application program. Typically, to Kerberize an application, you must have the application's source code.
You might recognize this problem as similar to making a service recognize PAM, and you'd be right. In fact, because PAM offers several Kerberos modules, any service that can use PAM for authentication can, through PAM, use Kerberos. This is probably the most flexible and convenient way to Kerberize a service. See Chapter 4 for more information about PAM.
Kerberos is designed for an environment in which there is one user per workstation. Because of the difficulty of sharing data between different processes running on the same Unix computer, Kerberos keeps tickets in the /tmp directory. If a user is sharing the computer with several other people, it is possible that the user's tickets can be stolen (copied by an attacker). Stolen tickets can then be used to obtain fraudulent service.
By design, Kerberos requires that there be a secure central server that maintains the master password database. To ensure security, a site should use the Kerberos Server for absolutely nothing beyond running the Kerberos Server program. The Kerberos Server must be kept under lock and key in a physically secure area. In some environments, maintaining such a server is an administrative and/or financial burden.
If the Kerberos Server goes down and is not replicated, the Kerberos network is unusable.
Adding to the difficulty of running a secure server is the fact that the Kerberos Server stores all passwords encrypted with the server's master key, which happens to be located on the same hard disk as the encrypted passwords. This means that, in the event that the Kerberos Server or its backups are compromised, all user passwords must be changed.
Kerberos does not have the computer authenticate itself to the user?that is, there is no way for a user sitting at a computer to determine whether the computer has been compromised. This failing is easily exploited by a knowledgable attacker.
 In fact, Trojan horses were a continuous problem at MIT's Project Athena.
For example, an intruder can modify the workstation's system software so every username/password combination typed is recorded automatically or sent electronically to another machine controlled by the attacker. Alternatively, a malicious attacker can simply modify the workstation's software to spuriously delete the user's files after the user has logged in and authenticated himself to the File Server Service. Both of these problems are consequences of the fact that, even in a networked environment, many workstations contain local copies of the programs that they run.
Another problem with Kerberos is that if a server password or a user password is broken or otherwise disclosed, it is possible for an eavesdropper to use that password to decrypt other tickets and use this information to spoof servers and users.
Kerberos is a workable system for network security, and it is widely used. But more importantly, the principles behind Kerberos are increasingly available in network security systems that are available directly from vendors.