Kerberos: Designing an Authentication System: a Dialogue in Four Scenes

Stolen from here and saved on my site so I can get at it again, without much searching.

Designing an Authentication System: a Dialogue in Four

Copyright 1988, 1997 Massachusetts Institute of Technology.
All Rights Reserved.

Originally written by Bill Bryant, February 1988.

Cleaned up and converted to HTML by Theodore Ts'o, February, 1997.
An afterword describing the changes in
Version 5 of the Kerberos protocol was also added.


This dialogue provides a fictitious account of the design of an
open-network authentication system called "Charon." As the dialogue
progresses, the characters Athena and Euripides discover the problems
of security inherent in an open network environment. Each problem
must be addressed in the design of Charon, and the design evolves
accordingly. Athena and Euripides don't complete their work until the
dialogue's close.

When they finish designing the system, Athena changes the system's
name to "Kerberos," the
name, coincidentally enough, of the authentication system that was
designed and implemented at MIT's Project Athena. The dialogue's
"Kerberos" system bears a striking resemblence to the system described
Kerberos: An Authentication Service for Open Network
presented at the Winter USENIX 1988, at Dallas,


Dramatis Personae


an up and coming system developer.

a seasoned developer and resident crank.

Scene I

A cubicle area.
Athena and Euripides are working at neighboring terminals.


Hey Rip, this timesharing system is a drag. I can't
get any work done because everyone else is logged in.

Don't complain to me. I only work here.

You know what we need? We need to give
everyone their
own workstation so they don't have to worry about sharing
computer cycles. And we'll use a network to connect all the
workstations so folks can communicate with one another.

Fine. So what do we need, about a
thousand workstations?

More or less.

Have you seen the size of a typical workstation's
disk drive? There isn't enough room for all the software
that you have on a timesharing machine.

I figured that out already. We can keep copies of the
system software on various server machines. When you login to
a workstation, the workstation accesses the system software by
making a network connection with one of the servers. This setup
lets a whole bunch of workstations use the same copy of the system
software, and it makes software updates convenient. You don't
have to trundle around to each workstation. Just modify the
system software servers.

All right. What are you going to do about personal
files? With a timesharing system I can login and get to
my files from any terminal that is connected to the system.
Will I be able to walk up to any workstation and automatically
get to my files? Or do I have to make like a PC user and keep
my files on diskette? I hope not.

I think we can use other machines to provide personal
file storage. You can login to any workstation and get to
your files.

What about printing? Does every workstation have
its own printer? Whose money are you spending anyway?
And what about electronic mail? How are you going to
distribute mail to all these workstations?

Ah . . . Well obviously we don't have the
cash to give
everyone a printer, but we could have machines
dedicated to print service. You send a job to a print server,
and it prints it for you. You could do sort of the same
thing with mail. Have a machine dedicated to mail service.
You want your mail, you contact the mail server and pick up
your mail.

Your workstation system sounds really good Tina.
When I get mine, you know what I'm going to do? I'm going
to find out your username, and get my workstation to think
that I am you. Then I'm going to contact the mail server
and pick up your mail. I'm going to contact your file server
and remove your files, and–

Can you do that?

Sure! How are these network servers going to know
that I'm not you?

Gee, I don't know. I guess I need to do some thinking.

Sounds like it. Let me know when you figure it out.

Scene II

Euripides' office, the next morning. Euripides sits at his desk,
reading his mail. Athena knocks on the door.


Well I've figured out how to secure an
open network environment so that unscrupulous folks like you cannot
use network services in other people's names.

Is that so? Have a seat.
She does.

Before I describe it, can I lay down one
ground rule about this discussion?

What's your rule?

Well suppose I say something like
the following: "I want my electronic mail, so I contact the mail
server and ask it to send the mail to my workstation." In reality
I'm not the entity that contacts the mail server. I'm using a
program to contact the mail server and retrieve my mail, a program
that is a CLIENT of the mail service program.

But I don't want to say "the client does such-and-such" every time
I refer to a transaction between the user and a network server.
I'd just as soon say "I do such-and-such," keeping in mind of
course that a client program is doing things on my behalf. Is that
okay with you?


Sure. No problem.

Good. All right, I'll begin by
stating the problem I have solved. In an open network environment,
machines that provide services must be able to confirm the
identities of people who request service. If I contact the mail
server and ask for my mail, the service program must be able to
verify that I am who I claim to be, right?


You could solve the problem
clumsily by requiring the mail server to ask for a password before I
could use it. I prove who I am to the server by giving it my

That's clumsy all right.
In a system like that, every server has to know your password. If
the network has one thousand users, each server has to know one
thousand passwords. If you want to change your password, you have
to contact all servers and notify them of the change. I take it
your system isn't this stupid.

My system isn't stupid. It works like this:
Not only do people have passwords, services
have passwords too. Each user knows her or his password,
each service program knows its password, and there's an
AUTHENTICATION SERVICE that knows ALL passwords–each user's password,
and each service's password. The authentication service stores
the passwords in a single, centralized database.

Do you have a name for this
authentication service?

I haven't thought of one yet. Do you have
any ideas?

What's the name of that fellow who
ferries the dead across the River Styx?


Yeah, that's him. He won't take you
across the river unless you can prove your identity.

There you go Rip, trying to rewrite
Greek mythology again. Charon doesn't care about your identity. He
just wants to make sure that you're dead.

Have you got a better name?

No, not really.

Then let's call the authentication
service "Charon."

Okay. I guess I should describe the system,

Let's say you want to use a service, the mail service.
In my system you cannot use a service unless, ah, Charon
tells the service that you are who you claim to be.
And you can't get the okay to use a service unless you have
authenticated yourself to Charon.
When you request authentication from Charon, you have to tell
Charon the service for which you want the okay.
If you want to use the mail server, you've got to tell Charon.

Charon asks you to prove your identity. You do so by providing
your secret password. Charon takes your password and compares
it to the one that is registered for you in the Charon database.
If the two passwords match, Charon considers your identity proven.

Charon now has to convince the mail server that you are
who you say you are. Since Charon knows all service passwords,
it knows the mail service's password. It's conceivable that
Charon could give you the password, which you could forward
to the mail service as proof that you have authenticated yourself
to Charon.

The problem is, Charon cannot give
you the password directly, because then you would know it.
The next time you wanted mail, you could
circumvent Charon and use the mail server without correctly
identifying yourself. You could even pretend to be someone else,
and use the mail server in that other person's name.

So instead of giving you the mail server's password, Charon
gives you a mail service TICKET. This ticket contains
a version of your username that has been ENCRYPTED USING the

Ticket in hand, you can now ask the mail service for your
mail. You make your request by telling the mail server
who you are, and furnishing the ticket that proves you are
who you say you are.

The server uses its password to decrypt the ticket,
and if the ticket decrypts properly, the server ends up
with the username that Charon placed in the ticket.

The service compares this name with the name you sent
along with the ticket. If the names match,
the mail server considers your identity proven and proceeds
to give you your mail.

What do you think of those apples?


I've got some questions.

I figured. Well go ahead.

When a service program decrypts a ticket,
how does it know that it has decrypted the ticket properly?

I don't know.

Maybe you should include the service's
name in the ticket. That way when a service decrypts a ticket, it
can gauge its success on whether or not it can find its name in the
decrypted ticket.

That sounds good to me. So the ticket
looks something like this:

(She scrawls the following on a pad of paper:)

TICKET -  {username:servicename}

So the service ticket contains just
your username and the servicename?

Encrypted with the service's password.

I don't think that's enough information
to make the ticket secure.

What do you mean?

Let's suppose you ask Charon
for a mail server ticket.
Charon prepares that ticket so that it has your username "tina"
in it. Suppose I copy that ticket as it wizzes by on its way
across the network from Charon to you. Suppose I convince my
insecure workstation that my username is "tina." The mail client
program on my workstation thinks I am you. In your name, the
program forwards the stolen ticket to the mail server. The server
decrypts the ticket and sees that it is valid. The username in
the ticket matches the name of the user who sent the ticket. The
mail server gives me your mail . . .

Oh! Well that's not so good.

But I think I know a way to fix
this problem. Or to at least provide a partial fix to it. I think
Charon should include more information in the service tickets it
produces. In addition to the username, the ticket should also
include the NETWORK ADDRESS from which the user asked Charon for the
ticket. That gives you an additional level of security.

I'll illustrate. Suppose I steal your mail ticket now.
The ticket has your workstation's network address in it,
and this address does not match my workstation's address.
In your name I forward the purloined ticket to the mail server.
The server program extracts the username and network address
from the ticket and attempts to match that information against
the username and network address of the entity that sent the
ticket. The username matches, but the network address does not.
The server rejects the ticket because obviously it was stolen.


Bravo, bravo! I wish I had thought of that.

Well that's what I'm around for.

So the revised ticket design looks like this:

She scrawls the following on a chalkboard:

TICKET -   {username:ws_address:servicename}

Now I'm really excited. Let's build a
Charon system and see if it works!

Not so fast. I have some
other questions about your system.

All right. (Athena leans forward in
her chair)

Sounds like I've got to get a
new ticket every time I want to use a service. If I'm putting in a
full day's work, I'll probably want to get my mail more than once.
Do I have to get a new ticket every time I want to get my mail? If
that's true, I don't like your system.

Ah . . . Well I don't see why tickets
can't be reusable. If you get a ticket for the mail server, you
ought to be able to use it again and again. For instance, when the
mail client program makes a request for service in your name, it
forwards a COPY of the ticket to the mail server.

That's better. But I still have problems.
You seem to imply that I have to give Charon my password
every time I want to use a service for which I don't have
a ticket. I login and want to access my files. I fire off
a request to Charon for the proper ticket and this means
that I've had to use my password. Then I want to read my mail.
Another request to Charon, I have to enter my password again.
Now suppose I want to send one of my mail messages to the print
server. Another Charon request and, well you get the picture.

Uh, yeah, I do.

And if that weren't bad enough, consider this:
it sounds like when you authenticate yourself to Charon,
you send your secret password over the network in cleartext.
Clever people like yours truly can monitor the
network and steal copies of people's passwords. If I've got your
password, I can use any service in your name.
Athena sighs.

These are serious problems.
Guess I need to go back to the drawing board.

Scene III

The next morning, Athena catches Euripides at the coffee area.
She taps him on the shoulder as he fills his cup.

The two head for the coffee machine.


I've got a new version of Charon that
solves our problems.

Really? That was quick.

Well, you know, problems of this nature
keep me up all night.

Must be your guilty conscience. Shall
we repair to yon small conference room?

Why not?
The two move to the small conference room.

I'll begin by stating the problems again,
but I'll invert them so that they become requirements of the system.
Athena clears her throat.

The first requirement: Users only have
to enter their passwords once, at the beginning of their workstation
sessions. This requirement implies that you shouldn't have to enter
your password every time you need a new service ticket. The second
requirement: passwords should not be sent over the network in clear



I'll start with the first requirement:
you should only have to use your password once. I've met this
requirement by inventing a new network service. It's called
the "ticket-granting" service, a service that issues Charon
tickets to users who have already proven their identity
to Charon. You can use this ticket-granting service if you
have a ticket for it, a ticket-granting ticket.

The ticket-granting service is really just a version of Charon
in as much as it has access to the Charon database.
It's a part of Charon that lets you authenticate yourself
with a ticket instead of a password.

Anyhow, the authentication system now works as follows: you login
to a workstation and use a program called kinit to contact the
Charon server. You prove your identity to Charon,
and the kinit program gets you a ticket-granting ticket.

Now say you want to get your mail from the mail server.
You don't have a mail server ticket yet, so you use the
"ticket-granting" ticket to get the mail server ticket for you.
You don't have to use your password to get the new ticket.


Do I have to get a new
"ticket-granting" ticket every time I need to get to another network

No. Remember, we agreed last time that
tickets can be reused. Once you have acquired a ticket-granting
ticket, you don't need to get another. You use the ticket-granting
ticket to get the other tickets you need.

Okay, that makes sense. And since you
can reuse tickets, once the ticket-granting service has given you a
ticket for a particular service, you don't need to get that
particular ticket again.

Yeah, isn't that elegant?

Okay, I buy it so far . . .
As long as you didn't have to send your password in cleartext over
the network when you got the ticket-granting ticket.

Like I said, I've solved that problem as well.
The thing is, when I say you have to contact Charon
to get the ticket-granting ticket, I make it sound as though
you have to send your password in cleartext over the network
to the Charon Server. But it doesn't have to be that way.

Here's really what happens. When you use the kinit
program to get the ticket-granting ticket, kinit doesn't
send your password to the Charon server, kinit sends only
your username.



Charon uses the username to look
up your password. Next Charon builds a packet of data that contains
the ticket-granting ticket. Before it sends you the packet, Charon
uses your password to encrypt the packet's contents.

Your workstation receives the ticket packet. You enter your
password. Kinit attempts to decrypt the ticket with the
password you entered. If kinit succeeds, you have
successfully authenticated yourself to Charon. You now possess a
ticket-granting ticket, and that ticket can get you the other
tickets you require.

How's that for some fancy thinking?


I don't know . . . I'm trying
to think myself. You know, I think the parts of the system that you
just described work pretty well. Your system requires me to
authenticate myself only once. Thereafter Charon can issue me
service tickets without my being aware of it. Seamless, seamless in
that regard. But there's something about the design of the service
ticket that troubles me somehow. It has to do with the fact that
tickets are reusable. Now I agree that they have to be reusable,
but reusable tickets are, by their nature, very dangerous.

What do you mean?

Look at it this way. Suppose
you are using an insecure workstation. In the course of your login
session you acquire a mail service ticket, a printing service
ticket, and a file service ticket. Suppose you inadvertantly leave
these tickets on the workstation when you logout.

Now suppose I login to the workstation and find those tickets.
I'm feeling like causing trouble, so I make the workstation think
that I am you. Since the tickets are made out in your name, I can
use the mail client program to access your mail, I can use the file
service client to access and remove your files, and I can use the
printing command to run up huge bills on your account. All because
these tickets have been accidentally left lying around.

And nothing can keep me from copying these tickets to a place of
my own. I can continue to use them for all eternity.


But that's an easy fix.
We just write a program that destroys a user's tickets after each
login session. You can't use tickets that have been


Well obviously your system must
have a ticket-destroying program, but it's foolish to make users
rely on such a thing. You can't count on users to remember to
destroy their tickets every time they finish a workstation session.
And even if you rely upon your users to destroy their tickets,
consider the following scenario.

I've got a program that watches the network and copies service
tickets as they zip accross the network. Suppose I feel like
victimizing you. I wait for you to begin a workstation session, I
turn on my program and copy a bunch of your tickets.

I wait for you to finish your session, and eventually you logout
and leave. I fiddle with my workstation's network software and
change its address so that it matches the address of the workstation
you were using when you acquired the tickets I copied. I make my
workstation believe that I am you. I have your tickets, your
username, and the correct network address. I can REPLAY these
tickets and use services in your name.

It doesn't matter that you destroyed your tickets before you
ended your workstation session. The tickets I have stolen are valid
for as long as I care to use them, because your current ticket
design does not place a limit on the number of times you can reuse a
ticket, or on how long a ticket remains valid.


Oh I see what you're saying!
Tickets can't be valid forever because they would then constitute a
huge security risk. We have to restrict the length of time for
which a ticket can be used, perhaps give each ticket some kind of
expiration date.

Exactly. I think each ticket needs
to have two additional pieces of information: a lifespan that
indicates the length of time for which the ticket is valid, and a
timestamp that indicates the date and time at which Charon issued
the ticket. So a ticket would look something like this:

Euripides goes to the chalkboard and scrawls the

TICKET   {username:address:servicename:lifespan:timestamp}

Now when a service decrypts tickets,
it checks the ticket's username and address against the name and
address of the person sending the ticket, and it uses the timestamp
and lifespan information to see if the ticket has expired.

All right. What kind of lifetime
should the typical service ticket have?

I don't know. Probably the length
of a typical workstation session. Say eight hours.

So if I sit at my workstation for more
than eight hours, all my tickets expire. That includes my
ticket-granting ticket. So I have to reauthenticate myself to
Charon after eight hours.

That's not unreasonable is it?

I guess not. So we're settled —
tickets expire after eight hours. Now I've got a question for you.
Suppose I have copied YOUR tickets from the network–

(Eyes twinkling) Aw, Tina!
You wouldn't really do that would you?

This is just for the sake of argument.
I've copied your tickets. Now I wait for you to logout. Suppose
you have a doctor's appointment or a class to attend, so you end
your workstation session after a couple of hours. You are a smart
boots and have destroyed your copies of the tickets before logging

But I've stolen your tickets, and they are good for about six
hours. That gives me ample time to pillage your files
and print one thousand copies of whatever in your name.

See, the lifetime-timestamp business works fine in the
event that a ticket thief chooses to replay the ticket
after the ticket has expired. If the thief can replay
the ticket before that . . .


Uh, well . . . Of course you are right.

I think we have run into a major
problem. (She sighs.)

I guess that means you'll be busy tonight.
Want more coffee?

Why not.

Scene IV

The next morning in Euripides' office. Athena knocks on the


You've got rings under your eyes
this morning.

Well, you know. Another one of those
long nights.

Have you solved the replay problem?

I think so.

Have a seat.
She does.

As usual, I feel compelled to
restate the problem. Tickets are reusable within a limited
timespan, say eight hours. If someone steals your tickets and
chooses to replay them before they expire, we can't do anything to
stop them.

That's the problem.

We could beat the problem if we
designed the tickets so they couldn't be reusable.

But then you would have to get
a new ticket every time you wanted to use a network service.

Right. That is a clumsy solution at
best. (Pause.) Ah, how do I proceed with my argument? (She
ponders for a moment.)

All right, I'm going to restate the problem again, this time in
the form of a requirement. A network service must be able to prove
that the person using a ticket is the same person to whom that
ticket was issued.

Let me trace the authentication process again and see if I can
tease out an appropriate way to illustrate my solution to this

I want to use a certain network service. I access that service
by starting a client program on my workstation. The client sends
three things to the service machine– my name, my workstation's
network address, and the appropriate service ticket.

The ticket contains the name of the person it was issued to and
the address of the workstation that person was using when he or she
acquired the ticket. It also contains an expiration date in the
form of a lifespan and a timestamp. All this information has been
encrypted in the service's Charon password.

Our current authentication scheme relies on the following tests:

  • Can the service decrypt the ticket?
  • Has the ticket expired?
  • Do the name and workstation address specified
    in the ticket match the name and address of the person who sent
    the ticket?

What do these tests prove? The first test proves that the ticket
either did or did not come from Charon. If the ticket cannot be
decrypted, it did not come from the real Charon. The real Charon
would have encrypted the ticket with the service's password. Charon
and the service are the only two entities that know the service's
password. If the ticket decrypts successfully, the service knows
that it came from the real Charon. This test prevents folks from
building fake Charon tickets.

The second test checks the ticket's lifespan and timestamp. If
it has expired, the service rejects the ticket. This test stops
people from using old tickets, tickets that perhaps were stolen.

The third test checks the ticket-user's name and address against
the name and address of the person specified in the ticket. If the
test fails, the ticket-user has obtained (perhaps surreptitiously)
another person's ticket. The ticket is of course rejected.

If the names and addresses do match, what has the test proved?
Nothing. Scallywags can steal tickets from the network, change
their workstation addresses and usernames appropriately, and rifle
other folks resources. As I pointed out yesterday, tickets can be
replayed as long as they haven't expired. They can be replayed
because a service cannot determine that the person sending the
ticket is actually the ticket's legitimate owner.

The service cannot make this determination because it does not
share a secret with the user. Look at it this way. If I'm on watch
at Elsinore, you know, the castle in Hamlet, and you are
supposed to relieve me, I'm not supposed to let you take my place
unless you can provide the correct password. That's the case where
the two of us share a secret. And it's probably a secret that
someone else made up for everyone who stands on watch.

So I was thinking last night, why not have Charon make up a
password for the legitimate ticket-owner to share with the service?
Charon gives a copy of this session key to the service, and a
copy to the user. When the service receives a ticket from a user,
it can use the session key to test the user's identity.


Wait a second. How is Charon going to
give both parties the session key?

The ticket-owner gets the session key
as part of the reply from Charon. Like this:

She scrawls the following on a chalkboard:

CHARON REPLY  -  [sessionkey|ticket]

The service's copy of the session key comes inside
the ticket, and the service gets the key when it decrypts the ticket.
So the ticket looks like this:

TICKET -  {sessionkey:username:address:servicename:lifespan:timestamp}

When you want to get to a service, the client program you
start builds what I call an AUTHENTICATOR.
The authenticator contains your name and your workstation's
address. The client encrypts this information with the
session key, the copy of the session key you received when
you requested the ticket.

AUTHENTICATOR - {username:address} encrypted with session key

After building the authenticator, the client sends it and the ticket
to the service. The service cannot decrypt the authenticator yet
because it doesn't have the session key. That key is in the
ticket, so the service first decrypts the ticket.

After decrypting the ticket, the service ends up with the
the following information:

  • The ticket's lifespan and timestamp;
  • The ticket-owner's name;
  • The ticket-owner's network address;
  • The session key.

The service checks to see if the ticket has expired.
If all is well in that regard, the service next uses
the session key to decrypt the authenticator. If
the decryption proceeds without a hitch, the service
ends up with a username and a network address. The service
tests this information against the name and address found
in the ticket, AND the name and address of the person who
sent the ticket and authenticator.
If everything matches, the service has determined that the
ticket-sender is indeed the ticket's real owner.

Athena pauses, clears her throat, drinks some coffee.

I think the session key-authenticator business takes care
of the replay problem.


Maybe. But I wonder . . . To break this version of
the system, I must have the proper authenticator for the service.

No. You must have the authenticator
AND the ticket for the service. The authenticator is worthless
without the ticket because the service cannot decrypt the
authenticator without first having the appropriate session key, and
the service cannnot get the appropriate session key without first
decrypting the ticket.

Okay, I understand that,
but didn't you say that when a client program contacts the server,
it sends the ticket and matching authenticator together?

Yes, I guess I said that.

If that's what actually happens,
what prevents me from stealing the ticket and authenticator at the
same time? I'm sure I could write a program to do the job. If I've
got the ticket and its authenticator, I believe I can use the two as
long as the ticket has not expired. I just have to change my
workstation address and username appropriately. True?

(Biting her lip) True.
How dispiriting.

Wait, wait, wait! This isn't such a big deal.
Tickets are reusable as long as they haven't expired, but that
doesn't mean that authenticators have to be reusable. Suppose we
design the system so that authenticators can only be used once.
Does that buy us anything?

Well, it might. Let's see,
the client program builds the authenticator, then sends it with the
ticket to the service. You copy both ticket and authenticator as
they move from my workstation to the server. But the ticket and
authenticator arrive at the server before you can send your copies.
If the authenticator can only be used once, your copy of it is no
good, and you lose when you attempt to replay your ticket and

Well, that's a relief. So all we have to do is invent a way
to make the authenticator a one-time usable thing.


No problem. Let's just put a
lifespan and timestamp on them. Suppose each authenticator has a
lifespan of a couple of minutes. When you want to use a service,
your client program builds the authenticator, stamps it with the
current time, then sends it and the ticket to the server.

The server receives the ticket and authenticator and goes about
its business. When the server decrypts the authenticator, it checks
the authenticator's lifespan and timestamp. If the authenticator
hasn't expired, and everything else checks properly, the server
considers you authenticated.

Suppose I copied the athenticator and ticket as they crossed the
network. I have to change my workstation's network address and my
username, and I have to do this all in a couple of minutes. That's
a pretty tall order. In fact I don't think it's possible. Unless
. . .

Well, here's a potential problem. Suppose that instead of
copying the ticket and authenticator as they travel from your
workstation to the server, I copy original ticket packet that comes
from Charon, the packet you receive when you ask Charon to give you
a ticket.

This packet, as I recall, has two copies of the session key in
it: one for you and one for the service. The one for the service is
hidden in the ticket and I can't get to it, but what about the other
one, the one you use to build authenticators?

If I can get that copy of the session key, I can build my own
authenticators, and if I can build my own authenticators, I can
break the system.


That's something I thought about last
night, but then I traced the process of acquiring tickets and found
that it wasn't possible to steal authenticators that way.

You sit down at a workstation and use the kinit program to get
your ticket-granting ticket. Kinit asks for your username,
and after you enter it, kinit forwards the name to Charon.

Charon uses your name to look up your password, then proceeds
to build a ticket-granting ticket for you. As part of this process,
Charon creates a session key that you will share with the
ticket-granting service. Charon puts a copy of the session key
in the ticket-granting ticket, and puts your copy in the
the ticket packet that you are about to receive. But before
it sends you this packet, Charon encrypts the whole thing
with your password.

Charon sends the packet across the network. Someone can copy the
packet as it goes by, but they can't do anything with it because it
has been encrypted with your password. Specifically, no one can
steal the ticket-granting session key.

Kinit receives the ticket packet and prompts you for a
password, which you enter. If you enter the correct password,
kinit can decrypt the packet and give you your copy of the
session key.

Now that you've taken care of the kinit business, you
want to get your mail. You start the mail client program. This
program looks for a mail service ticket and doesn't find one (after
all, you haven't tried to get your mail yet). The client must use
the ticket-granting ticket to ask the ticket-granting service for a
mail service ticket.

The client builds an authenticator for the ticket-granting
transaction and encrypts the authenticator with your copy of the
ticket-granting session key. The client then sends Charon the
authenticator, the ticket-granting ticket, your name, your
workstation's address, and the name of the mail service.

The ticket-granting service receives this stuff and runs through
the authentication checks. If everything checks properly, the
ticket-granting service ends up with a copy of the session key that
it shares with you. Now the ticket-granting service builds you a
mail service ticket, and during this process, creates a new session
key for you to share with the mail service.

The ticket-granting service now prepares a ticket packet to send
back to your workstation. The packet contains the ticket and your
copy of the mail service session key. But before it sends the
packet, the ticket-granting service encrypts the packet with its
copy of the TICKET-GRANTING session key. That done, the packet is
sent on its way.

So here comes the mail service ticket packet, loping across the
network. Suppose some network ogre copies it as it goes by. The
ogre is out of luck because the packet is encrypted with the
ticket-granting session key; you and the ticket-granting service are
the only entities that know this key. Since the ogre cannot decrypt
the mail ticket packet, the ogre cannot discover the MAIL SESSION
KEY. Without this session key, the ogre cannot use any of the mail
service tickets you might subsequently send across the network.

So I think we're safe. What do you think?



Perhaps! Is that all you can say!

(laughing) Don't get upset.
You should know my ways by now. I guess it is mean of me, and you
up half the night.


All right, three-quarters of the night.
Actually, the system is beginning to sound acceptible. This session
key business solves a problem that I thought of last night: the
problem of mutual authentication.
Mind if I talk for a minute?

(A trifle coldly) Be my guest.

You are so kind.
(Euripides clears his throat.) Last night, while visions of
session keys and authenticators danced in your head, I was trying to
find new problems with the system, and I found one that I thought
was pretty serious. I'll illustrate it by way of the following

Suppose you are sick of your current job and have determined that
it is in your best interest to move on. You want to print your
resume on the company's wizz-bang laser printer so that headhunters
and potential employers can take note of your classiness.

So you enter the printing command, and direct it to send the
resume to the appropriate print server. The command gets the proper
service ticket, if you don't already have it, then sends the ticket
in your name to the appropriate print server. At least that's where
you think it's headed. You don't in fact know that the request is
headed for the right print server.

Suppose that some unscrupulous hacker–say it's your boss–has
screwed system around so that he redirects your request and its
ticket to the print server in his office. His print service program
doesn't care about the ticket or its contents. It throws away the
ticket and sends a message to your workstation indicating that the
ticket passed muster, and that the server is ready and willing to
print your job. The printing command sends the job to the
fraudulant print server and the enemy ends up with your resume.

I'll state the problem by way of contrast. Without session keys
and authenticators, Charon can protect its servers from false users,
but it cannot protect its users from false servers. The system
needs a way for client programs to authenticate the server before
sending sensitive information to the service. The system must allow
for mutual authentication.

But the session key solves this problem as long as you design
your client programs properly. Back to the print server scenario.
I want a print client program that makes sure the service it's
sending jobs to is the legitimate service.

Here's what such a program does. I enter the printing command
and give it a filename, the name of my resume. Assume that I have a
print service ticket and session key. The client program uses the
session key to build an authenticator, then sends the authenticator
and ticket to the "supposed" print server. The client DOES NOT send
the resume yet; it waits for a response from the service.

The real service receives the ticket and authenticator, decrypts
the ticket and extracts the session key, then uses the session key
to decrypt the authenticator. This done, the service runs all the
appropriate authentication tests.

Assume the tests confirm my identity. Now the server prepares a
reply packet so that it can prove its identity to the client
program. It uses its copy of the session key to encrypt the reply
packet, then sends the packet to the waiting client.

The client receives the packet and attempts to decrypt it with my
copy of the session key. If the packet decrypts properly and yields
the correct server response message, my client program knows that
the server that encrypted the packet is the real server. Now the
client sends the resume job to the print service.

Suppose my boss screwed around the system so that his print
server poses as the one I want. My client sends the authenticator
and ticket to the "print service" and waits for a response. The
fake print service cannot generate the correct response because it
cannot decrypt the ticket and get the session key. My client will
not send the job unless it receives the correct response.
Eventually the client gives up waiting and exits. My print job does
not get completed, but at least my resume did not end up on the desk
of the enemy.

You know, I think we have a solid basis on which to implement the
Charon Authentication System.


Perhaps. Anyway, I don't
like the name "Charon."

You don't? Since when?

I've never liked it, because the name
doesn't make sense. I was talking to my Uncle Hades about it the
other day, and he suggested another name, the name of his
three-headed watch dog.

Oh, you mean "Cerberus."

Bite your tongue Rip! "Cerberus"
indeed . . .

Er, isn't that the name?

Yeah, if you happen to be a Roman!
I'm a Greek goddess, he's a Greek watch dog, and his name is
"Kerberos," "Kerberos" with a K.

Okay, okay, don't throw thunderbolts.
I'll buy the name. Actually, it has a nice ring to it. Adios
Charon and hello to Kerberos.


The dialogue was written in 1988 to help its readers understand the
fundamental reasons for why the Kerberos V4 protocol was the way it
was. Over the years, it has served this job very well.

When I converted this document to HTML, I was amazed how much of
this document was still applicable for the Kerberos V5 protocol.
Although many things were changed, the basic core ideas of the
protocol have remained the same. Indeed, there are only two changes
where Kerberos V5 differs from description of the "Kerberos"
protocol in this dialoge.

The first change was born out of the recognition that using a small
five minute time skew wasn't necessarily sufficient to prevent
replay attacks from an attacker who used a program to automatically
grab the ticket and the authenticator as they traversed the network,
and then immediately resent them to launch a replay attack.

In Kerberos V5, authenticators are made to be truly "once-only" by
having servers which accept tickets to have a "replay cache" which
keeps note of authenticators have been presented to the server
recently. If an attacker tries to snatch an authenticator and reuse
it, even during the five-minute acceptance window, the replay cache
will be able to determine that the authenticator has already been
presented to the server.

The second major change to the protocol is that the ticket is no
longer encrypted in the user's password when it is sent from the
Kerberos server to kinit during the initial ticket
exchange. The ticket is already encrypted in the ticket granting
server's secret key; furthermore when it is actually used to obtain
other tickets, it gets sent in the network in the clear anyway.
Hence, there is no reason why the ticket should be encrypted again
in the user's password. (The rest of the Kerberos server's reply to
the user, containing for example the user's copy of the ticket
session key, is still encrypted in the user's password, of

A similar change was also made to the ticket granting service
(TGS) protocol; tickets returned by TGS are also no longer
encrypted by the ticket-granting ticket's service key, since
application tickets are already encrypted by the application
server's secret key. So for example, the packet that in Kerberos V4
which would have looked like this:

KDC_REPLY = {TICKET, client, server, K_session}K_user

where "{X}K_Y" is read "X encrypted using key K_Y" and

TICKET = {client, server, start_time, lifetime, K_session}K_server

In Kerberos V5, the KDC_REPLY now would look like this:

KDC_REPLY = TICKET, {client, server, K_session}K_user

Of course, there are many new features in Kerberos V5 as well. Users
can now securely forward their tickets so that they can be used at
another network location; in addition, users may also delagate a
subset of their authorization rights to a server, so that the server
can act as a proxy on behalf of a user. Other new features include
the ability to replace DES with a more secure cryptographic
algorithm, such as triple-DES. Readers who are interested in more
of the changes between Kerberos V4 and V5 are invited to read The
Evolution of the Kerberos Authentication System
, which was
authored by
Cliff Neumann and
Theodore Ts'o.

I hope you've enjoyed this quick little introduction to the
Kerberos protocol. I wish you well in your futher explorations!

Theodore Ts'o, February 1997.

Permission to use, copy, modify, and
distribute this documentation for any purpose and
without fee is hereby granted, provided that the above copyright
notice appear in all copies and that both that copyright notice and
this permission notice appear in supporting documentation, and that
the name of M.I.T. not be used in advertising or publicity pertaining
to distribution of the documentation without specific, written prior
permission. M.I.T. makes no representations about the suitability of
this documentation for any purpose. It is provided "as is" without express
or implied warranty.

For comments/suggestions about this page, mail:

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s