tinc/doc/SECURITY

This is the security documentation for tinc, a Virtual Private Network daemon.

   Copyright 2000 Guus Sliepen <guus@sliepen.warande.net>,
             2000 Ivo Timmmermans <itimmermans@bigfoot.com>

   Permission is granted to make and distribute verbatim copies of
   this documentation provided the copyright notice and this
   permission notice are preserved on all copies.

   Permission is granted to copy and distribute modified versions of
   this documentation under the conditions for verbatim copying,
   provided that the entire resulting derived work is distributed
   under the terms of a permission notice identical to this one.

   $Id: SECURITY,v 1.1.2.3 2000/09/25 20:08:50 guus Exp $


1.  Authentication
------------------

The authentication protocol (see protocol.c for the up-to-date version) is:

   Client               Server
   send_id(u)
                        send_challenge(R)
   send_chal_reply(H)                   
                        send_id(u)
   send_challenge(R)
                        send_chal_reply(H)
   ---------------------------------------
   Any negotations about the meta protocol
   encryption go here(u).
   ---------------------------------------
   send_ack(u)
                        send_ack(u)
   ---------------------------------------
   Other requests(E)...

   (u) Unencrypted,
   (R) RSA,
   (H) SHA1,
   (E) Encrypted with symmetric cipher.

See section 4 for a detailed example version of the authentication.

Authentication in tinc will be done in a way that is very similar to the way
the SSH (Secure SHell) authentication protocol works. It is based on public
key cryptography.

Every tinc host has its own public/private key pair. Suppose there are two
tinc hosts, A and B. If A and B trust each other, they store a copy of
eachothers public key (in the same way passphrases were stored in versions
of tinc <= 1.0pre2). They know these public keys beforehand, and the origin
of the public keys has to be known for sure.

To make sure that when a connection is made from A to B that B knows A is
really who he claims to be, B encrypts a totally random string of bytes with
A's public key. B also calculates the hash value from the unencrypted random
string. B then sends the encrypted string to A. A then has to decrypt the
string, calculate the hash value from that string and send it back to B. Since
only he who possesses A's private key can decrypt this string, only he can send
back the correct hash value. So, if B receives the same hash value he
calculated himself, he knows for sure A is A.

Both SSH and tinc use RSA for the public key cryptography. SSH uses MD5 as a
secure hash algorithm, tinc uses SHA1. The reason for our choice of SHA1 is
the fact that SHA1 is 160 bits instead of 128 (MD5), which makes brute force
attacks harder. Also, the OpenSSL documentation recommends SHA1.

2.  Key exchange
----------------

The rest of the meta connection in tinc will be encrypted with a symmetric
block cipher, since RSA is not really suited for this. When a connection is
made, both sides have to agree on a key for this block cipher. To make sure
that this key exchange is also done securely, and no man-in-the-middle attack
is possible, RSA would be the best choice for exchanging keys.

Instead of doing RSA encryption again, tinc will use a part of the random
string that was exchanged during the authentication phase as the key for the
symmetric cipher. Some symmetric ciphers require a random initialisation vector
for improved security. This vector can be taken from the random string as well.

Is this secure? I (Guus Sliepen) think at this moment that it is:

- Since the random string cannot be decrypted by anyone eavesdropping or
  playing man-in-the-middle, the symmetric key cannot be known by sniffing.
- The unencrypted returned hash value is supposed to be cryptographically
  secure. Furthermore, it can only at most give a way 160 bits of information
  from the complete random string which is longer than the key for the
  symmetric cipher, so very few bits will actualy contain information about
  the symmetric cipher key alone, if any.
- If the RSA encryption is cracked, the rest of the communications can be
  decrypted anyway.
- If the symmetric cipher encryption is cracked without using the information
  from the encrypted random strings or the hash values, this still won't give
  the full plaintext for the random string, so it won't facilitate a known-
  plaintext attack on the RSA encryption.
- RSA and symmetric ciphers are fundamentally different. It is very unlikely
  that the overlap of both will create any interference that will facilitate
  an easier-than-brute-force attack.

Other options for key exchange could be:

* A second exchange of RSA encrypted random strings.
  This is equal to the former scheme just without knowing the hash value of
  the unecrypted random string. Information theory tells that two seperate
  RSA messages are as secure as one if the total amount of bits sent is the
  same, so enlarging the challenge will make one exchange just as secure as
  two seperate exchanges.
  
* Diffie-Hellman with RSA signing.
  This should be very secure, but there are a lot of pitfalls with using both
  encryption with public keys and private keys together with the same keypair.

* Diffie-Hellman with passphrases.
  This is what tinc <= 1.0pre2 used to do. Passphrases are secret, exchanging
  them must be done with great care, nobody may eavesdrop. Exchanging public
  keys on the other hand is much safer, everybody may eavesdrop, just as long
  as you are sure that the public key itself belongs to the right owner.

3.  Symmetric cipher
--------------------

Since the generalized encryption functions of OpenSSL are used, any symmetric
cipher that is available in OpenSSL could possibly be used. The default however
will be Blowfish. Blowfish is widely in use and still has not been cracked
today (as far as we know). It also is one of the faster ciphers available.

4.  Detailed "example" of communication
---------------------------------------

Tinc uses a peer-to-peer protocol, but during the authentication phase we will
make a distinction between a server (a tinc daemon listening for incoming
connections) and a client (a tinc daemon that is trying to connect to the tinc
daemon playing server).

The message strings here are kept short for clarity. The real length of the
exchanged messages is indicated. The capital words ID, CHALLENGE, CHAL_REPLY
and ACK are in reality replaced by the numbers 1, 2, 3 and 4 respectively.

daemon	message
--------------------------------------------------------------------------
server	<listening for connection>
client  <tries to connect>
server  <accepts connection>
client	ID client 8 0
              |   | +-> options
              |   +---> version
              +-------> name of tinc daemon
server  CHALLENGE 57fb4b2ccd70d6bb35a64c142f47e61d
                  \________/\__/
                       |      +----> 64 bits initial vector and
                       +-----------> 448 bits symmetric cipher key for meta
                                     data sent to the server
                  \______________________________/
                                 +-> 2048 bits totally random string, encrypted
                                     with client's public RSA key
client	CHAL_REPLY 191e23
                      +-> 160 bits SHA1 value of the complete decrypted
                          CHALLENGE sent by the server
server	ID server 8 0
              |   | +-> options
              |   +---> version
              +-------> name of tinc daemon
client	CHALLENGE da02add1817c1920989ba6ae2a49cecb
                  \________/\__/
                       |      +----> 64 bits initial vector and
                       +-----------> 448 bits symmetric cipher key for meta
                                     data sent to the client
                  \______________________________/
                                 +-> 2048 bits totally random string, encrypted
                                     with server's public RSA key
server	CHAL_REPLY 2bdeed
                      +-> 160 bits SHA1 value of the complete decrypted
                          CHALLENGE sent by the client
client	ACK
server	ACK
--------------------------------------------------------------------------

When the server receives the ACK from the client, it should prepare itself
for the fact that any subsequent data will be encrypted with the key the server
sent itself in the CHALLENGE. Ofcourse, this key is taken from the decrypted
version of that CHALLENGE, so that we will know for sure only the real client
can send us messages. The same goes for the client when it receives an ACK.