Network Working Group M. Chatel
Request for Comments: 1919 Consultant
Category: Informational March 1996
Classical versus Transparent IP Proxies
Status of this Memo
This memo provides information for the Internet community. This memo
does not specify an Internet standard of any kind. Distribution of
this memo is unlimited.
Abstract
Many modern IP security systems (also called "firewalls" in the
trade) make use of proxy technology to achieve Access control. This
document eXPlains "classical" and "transparent" proxy techniques and
attempts to provide rules to help determine when each proxy system
may be used without causing problems.
Table of Contents
1. Background . . . . . . . . . . . . . . . . . . . . . . . . . 2
2. Direct communication (without a proxy) . . . . . . . . . . . 3
2.1. Direct connection example . . . . . . . . . . . . . . . . 3
2.2. Requirements of direct communication . . . . . . . . . . . 5
3. Classical application proxies . . . . . . . . . . . . . . 5
3.1. Classical proxy session example . . . . . . . . . . . . . 6
3.2. Characteristics of classical proxy configurations . . . 12
3.2.1. IP addressing and routing requirements . . . . . . . . 12
3.2.2. IP address hiding . . . . . . . . . . . . . . . . . . 14
3.2.3. DNS requirements . . . . . . . . . . . . . . . . . . . 14
3.2.4. Software requirements . . . . . . . . . . . . . . . . 15
3.2.5. Impact of a classical proxy on packet filtering . . . 15
3.2.6. Interconnection of conflicting IP networks . . . . . . 16
4. Transparent application proxies . . . . . . . . . . . . . 19
4.1. Transparent proxy connection example . . . . . . . . . . 20
4.2. Characteristics of transparent proxy configurations . . 26
4.2.1. IP addressing and routing requirements . . . . . . . . 26
4.2.2. IP address hiding . . . . . . . . . . . . . . . . . . 28
4.2.3. DNS requirements . . . . . . . . . . . . . . . . . . . 28
4.2.4. Software requirements . . . . . . . . . . . . . . . . 29
4.2.5. Impact of a transparent proxy on packet filtering . . 30
4.2.6. Interconnection of conflicting IP networks . . . . . . 31
5. Comparison chart of classical and transparent proxies . . 31
6. Improving transparent proxies . . . . . . . . . . . . . . 32
7. Security Considerations . . . . . . . . . . . . . . . . . 34
8. Acknowledgements . . . . . . . . . . . . . . . . . . . . . 34
9. References . . . . . . . . . . . . . . . . . . . . . . . . 35
1. Background
An increasing number of organizations use IP security systems to
provide specific access control when crossing network security
perimeters. These systems are often deployed at the network boundary
between two organizations (which may be part of the same "official"
entity), or between an organization"s network and a large public
internetwork sUCh as the Internet.
Some people believe that IP firewalls will become commodity products.
Others believe that the introduction of IPv6 and of its improved
security capabilities will gradually make firewalls look like stopgap
solutions, and therefore irrelevant to the computer networking scene.
In any case, it is currently important to examine the impact of
inserting (and removing) a firewall at a network boundary, and to
verify whether specific types of firewall technologies may have
different effects on typical small and large IP networks.
Current firewall designs usually rely on packet filtering, proxy
technology, or a combination of both. Packet filtering (although hard
to configure correctly in a security sense) is now a well documented
technology whose strengths and weaknesses are reasonably understood.
Proxy technology, on the other hand, has been deployed a lot but
studied little. Furthermore, many recent firewall products support a
capability called "transparent proxying". This type of feature has
been subject to much more marketing attention than actual technical
analysis by the networking community.
It must be remembered that the Internet"s growth and success is
strongly related to its "open" nature. An Internet which would have
been segmented from the start with firewalls, packet filters, and
proxies may not have become what it is today. This type of discussion
is, however, outside the scope of this document, which just attempts
to provide an understandable description of what are network proxies,
and of what are the differences, strengths, and weaknesses of
"classical" and "transparent" network proxies. Within the context of
this document, a "classical" proxy is the older (some would say old-
fashioned) type of proxy of the two.
Also note that in this document, the Word "connection" is used for an
application session that uses TCP, while the word "session" refers to
an application dialog that may use UDP or TCP.
2. Direct communication (without a proxy)
In the "normal" Internet world, systems do not use proxies and simply
use normal TCP/IP to communicate with each other. It is important
(for readers who may not be familiar with this) to take a quick look
at the operations involved, in order to better understand what is the
exact use of a proxy.
2.1 Direct connection example
Let"s take a familiar network session and describe some details of
its operation. We will look at what happens when a user on a
client system "c.dmn1.com" sets up an FTP connection to the server
system "s.dmn2.com". The client system"s IP address is
c1.c2.c3.c4, the server"s IP address is s1.s2.s3.s4.
+---------------+ +----------+ +---------------+
/ IP
c.dmn1.com ----+ network(s) +---- s.dmn2.com
(c1.c2.c3.c4) / (s1.s2.s3.s4)
+---------------+ +----------+ +---------------+
The user starts an instance of an FTP client program on the client
system "c.dmn1.com", and specifies that the target system is
"s.dmn2.com". On command-line systems, the user typically types:
ftp s.dmn2.com
The client system needs to convert the server"s name to an IP
address (if the user directly specified the server by address,
this step is not needed).
Converting the server name to an IP address requires work to be
performed which ranges between two extremes:
a) the client system has this name in its hosts file, or has
local DNS caching capability and successfully retrieves the
name of the server system in its cache. No network activity
is performed to convert the name to an IP address.
b) the client system, in combination with DNS name servers,
generate DNS queries that eventually propagate close to the
root of the DNS tree and back down the server"s DNS branch.
Eventually, a DNS server which is authoritative for the
server system"s domain is queried and returns the IP
address associated with "s.dmn2.com" (depending on the case,
it may return this to the client system directly or to an
intermediate name server). Ultimately, the client system
oBTains a valid IP address for s.dmn2.com. For simplicity,
we assume the server has only one IP address.
+---------------+ +--------+ +---------------+
/ IP
c.dmn1.com ---+ network(s) +--- s.dmn2.com
(c1.c2.c3.c4) / (s1.s2.s3.s4)
+---------------+ +--------+ +---------------+
A /
address for /
s.dmn2.com? /
/
/
+--------+ s.dmn2.com? +--------+
+----> DNS -------------> DNS
server server
+-------- X <------------- Y
s1.s2.s3.s4 +--------+ s1.s2.s3.s4 +--------+
Once the client system knows the IP address of the server system,
it attempts to establish a connection to the standard FTP
"control" TCP port on the server (port 21). For this to work, the
client system must have a valid route to the server"s IP address,
and the server system must have a valid route to the client"s IP
address. All intermediate devices that behave like IP gateways
must have valid routes for both the client and the server. If
these devices perform packet filtering, they must ALL allow the
specific type of traffic required between C and S for this
specific application.
+---------------+ +---------------+
c.dmn1.com s.dmn2.com
(c1.c2.c3.c4) (s1.s2.s3.s4)
+---------------+ +---------------+

route to S route to C
V V

A A
route to C route to S

C S C
+----+ <-- +----+ --> +----+ <-- +----+
G1 -------- Gx -------- Gy --------- Gn
+----+ --> +----+ <-- +----+ --> +----+
S C S
The actual application work for the FTP session between the client
and server is done with a bidirectional flow of TCP packets
between the client"s and server"s IP addresses.
The FTP protocol uses a slightly complex protocol and TCP
connection model which is, luckily, not important to the present
discussion. This allows slightly shortening this document...
2.2 Requirements of direct communication
Based on the preceding discussion, it is possible to say that the
following is required for a direct session between a client and
server to be successful:
a) If the client uses the NAME of the server to reference it,
the client must either have a hardcoded name-to-address
binding for the server, or it must be able to resolve the
server name (typically using DNS). In the case of DNS, this
implies that the client and server must be part of the same
DNS architecture or tree.
b) The client and server must be part of the same internetwork:
the client must have a valid IP route towards the server,
the server must have a valid IP route towards the client,
and all intermediate IP gateways must have valid routes
towards the client and server ("IP gateway" is the RFC
standard terminology; people often use the term "IP router"
in computer rooms).
c) If there are devices on the path between the client and
server that perform packet filtering, all these devices must
permit the forwarding of packets between the IP address of
the client and the IP address of the server, at least for
packets that fit the protocol model of the FTP application
(TCP ports used, etc.).
3. Classical application proxies
A classical application proxy is a special program that knows one (or
more) specific application protocols. Most application protocols are
not symetric; one end is considered to be a "client", one end is a
"server".
A classical application proxy implements both the "client" and
"server" parts of an application protocol. In practice, it only needs
to implement enough of the client and server protocols to accomplish
the following:
a) accept client sessions and appear to them as a server;
b) receive from a client the name or address of the final target
server (this needs to be passed over the "client-proxy" session
in a way that is application-specific);
c) setup a session to the final server and appear to be a client
from the server"s point of view;
d) relay requests, responses, and data between the client and
server;
e) perform access controls according to the proxy"s design
criteria (the main goal of the proxy, after all).
The functional goal of the proxy is to relay application data between
clients and servers that may not have direct IP connectivity. The
security goal of the proxy is to do checks and types of access
controls that typical client and server software do not support or
implement.
The following information will make it clear that classical proxies
can offer many hidden benefits to the security-conscious network
designer, at the cost of deploying client software with proxy
capabilities or of educating the users on proxy use.
Client software issues are now easier to handle, given the increasing
number of popular client applications (for Web, FTP, etc.) that offer
proxy support. Designers developing new protocols are also more
likely to plan proxy capability from the outset, to ensure their
protocols can cross the many existing large corporate firewalls that
are based at least in part on classical proxy technology.
3.1 Classical proxy session example
We will repeat our little analysis of an FTP session. This time,
the FTP session is passing through a "classical" application proxy
system. As is often the case (although not required), we will
assume that the proxy system has two IP addresses, two network
interfaces, and two DNS names.
The proxy system is running a special program which knows how to
behave like an FTP client on one side, and like an FTP server on
the other side. This program is what people call the "proxy". We
will assume that the proxy program is listening to incoming
requests on the standard FTP control port (21/tcp), although this
is not always the case in practice.
+---------------+ +----------+
/ IP
c.dmn1.com ----+ network(s) +----------+
(c1.c2.c3.c4) /
+---------------+ +----------+ +-----------------+
(p1.p2.p3.p4)
proxy1.dmn3.com

proxy2.dmn4.com
(p5.p6.p7.p8)
+---------------+ +----------+ +-----------------+
/ IP
s.dmn2.com ----+ network(s) +----------+
(s1.s2.s3.s4) /
+---------------+ +----------+
The user starts an instance of an FTP client program on the client
system "c.dmn1.com", and MUST specify that the target system is
"proxy1.dmn3.com". On command-line systems, the user typically
types:
ftp proxy1.dmn3.com
The client system needs to convert the proxy"s name to an IP
address (if the user directly specified the proxy by address, this
step is not needed).
Converting the proxy name to an IP address requires work to be
performed which ranges between two extremes:
a) the client system has this name in its hosts file, or has
local DNS caching capability and successfully retrieves the
name of the proxy system in its cache. No network activity
is performed to convert the name to an IP address.
b) the client system, in combination with DNS name servers,
generate DNS queries that eventually propagate close to the
root of the DNS tree and back down the proxy"s DNS branch.
Eventually, a DNS server which is authoritative for the
proxy system"s domain is queried and returns the IP
address associated with "proxy1.dmn3.com" (depending on the
case, it may return this to the client system directly or
to an intermediate name server). Ultimately, the client
system obtains a valid IP address for proxy1.dmn3.com.
+---------------+ +--------+
/ IP
c.dmn1.com --------+ network(s) +------------+
(c1.c2.c3.c4) /
+---------------+ +--------+ +-----------------+
A / (p1.p2.p3.p4)
address for / proxy1.dmn3.com
proxy1.dmn3.com? / ...
/ +-----------------+
/
/
+--------+ proxy1.dmn3.com? +--------+
+--------> DNS ------------------> DNS
server server
+------------ X <------------------ Y
p1.p2.p3.p4 +--------+ p1.p2.p3.p4 +--------+
Once the client system knows the IP address of the proxy system,
it attempts to establish a connection to the standard FTP
"control" TCP port on the proxy (port 21). For this to work, the
client system must have a valid route to the proxy"s IP address,
and the proxy system must have a valid route to the client"s IP
address. All intermediate devices that behave like IP gateways
must have valid routes to both the client and the proxy. If these
devices perform packet filtering, they must ALL allow the specific
type of traffic required between C and P1 for this specific
application (FTP).
Finally, the proxy system must accept this incoming connection,
based on the client"s IP address (the purpose of the proxy is
generally to do access control, after all).
+---------------+ ...
c.dmn1.com proxy1.dmn3.com
(c1.c2.c3.c4) (p1.p2.p3.p4)
+---------------+ +-----------------+

route to P1 route to C
V V

A A
route to C route to P1

C P1 C
+----+ <-- +----+ --> +----+ <-- +----+
G1 -------- Gx -------- Gy --------- Gn
+----+ --> +----+ <-- +----+ --> +----+
P1 C P1
The actual application work for the FTP session between the client
and proxy is done with a bidirectional flow of TCP packets between
the client"s and proxy"s IP addresses.
For this to work, the proxy FTP application MUST fully support the
FTP protocol and look identical to an FTP server from the client"s
point of view.
Once the client<->proxy session is established, the final target
server name must be passed to the proxy, since, when using a
"classical" application proxy, a way MUST be defined for the proxy
to determine the final target system. This can be achieved in
three ways:
a) The client system supplies the name or address of the final
target system to the proxy in a method that is compatible
with the specific application protocol being used (in our
example, FTP). This is generally considered to be the main
problem with classical proxies, since for each application
being proxied, a method must be defined for passing the
name or address of the final target system. This method
must be compatible with every variant of client application
that implements the protocol (i.e. the target-passing
method must fit within the MINIMUM functionalities required
by the specific application protocol).
For the FTP protocol, the generally popular method for
passing the final server name to the proxy is as follows:
When the proxy prompts the FTP client for a username, the
client specifies a string of the form:
target_username@target_system_name
or
target_username@target_ip_address
The proxy will then know what is the final target system.
The target_username (and the password supplied by the
client) will be forwarded "as is" by the proxy to the final
target system.
A well-known example of an FTP proxy that behaves in this way
is the "ftp-gw" program which is part of the Trusted
Information System"s firewall toolkit, available by anonymous
FTP at ftp.tis.com. Several commercial firewalls also support
this de-facto standard.
b) If there is only one possible final destination, the proxy
may be configured to know this destination in advance.
Since the IP address of the client system is known when the
proxy must make this decision, the proxy can (if required)
select a different destination based on the IP address of
the client.
c) The client software may also support capabilities that allow
it to present to the user the illusion of a direct session
(the user just specifies the final target system, and the
client software automatically handles the problem of
reaching to the proxy system and passing the name or address
of the final target system in whatever mutually-acceptable
form).
A well-known example of a system that provides modified
client software, proxy software, and that provides the
illusion of transparency is NEC"s SOCKS system, available by
anonymous FTP at ftp.nec.com.
Alternatively, several FTP client applications support the
"username@destination_host" de-facto standard implemented
(for example) by the "ftp-gw" proxy application.
Once the FTP proxy application knows the name or IP address of the
target system, it can choose to do two things:
a) Setup a session to the final target system, the more
frequent case.
b) Decide (based on some internal configuration data) that it
cannot reach the final target system directly, but must go
through another proxy. This is rare today, but may become
temporarily common due to the current shortage of IP
network numbers which encourages organizations to deploy
"hidden" network numbers which are already assigned
elsewhere. Sessions between systems which have the same
IP network number but which belong to different actual
networks may require going through two proxy systems.
This is discussed in more detail in section 3.2.6,
"Interconnection of conflicting IP networks".
If the FTP proxy decides to connect directly to the target system,
and what it has is the target system name, it will need to convert
the target system name into an IP address. If this process
involves DNS resolution, something like the following will happen:
+-----------------+
proxy1.dmn3.com
(p1.p2.p3.p4) +--------+
/ IP
proxy2.dmn4.com --------+ network(s) +------------+
(p5.p6.p7.p8) /
+-----------------+ +--------+ +---------------+
A / (s1.s2.s3.s4)
address for / s.dmn2.com
s.dmn2.com? /
/ +---------------+
/
/
+--------+ s.dmn2.com? +--------+
+--------> DNS ------------------> DNS
server server
+------------ X <------------------ Y
s1.s2.s3.s4 +--------+ s1.s2.s3.s4 +--------+
Once the proxy system knows the IP address of the server system,
it attempts to establish a connection to the standard FTP
"control" TCP port on the server (port 21). For this to work, the
proxy system must have a valid route to the server"s IP address,
and the server system must have a valid route to at least one of
the proxy"s IP address. All intermediate devices that behave like
IP gateways must have valid routes to both the proxy and the
server. If these devices perform packet filtering, they must ALL
allow the specific type of traffic required between the proxy and
S for this specific application.
+-----------------+
proxy1.dmn3.com
(p1.p2.p3.p4)
+----------------+
proxy2.dmn4.com s.dmn2.com
(p5.p6.p7.p8) (s1.s2.s3.s4)
+-----------------+ +----------------+

route to S route to P2
V V

A A
route to P2 route to S

P2 S P2
+----+ <-- +----+ --> +----+ <-- +----+
G1 -------- Gx -------- Gy --------- Gn
+----+ --> +----+ <-- +----+ --> +----+
S P2 S
The actual FTP application work between the proxy and server is
done with a bidirectional flow of TCP packets between the proxy"s
and server"s IP addresses.
What actually happens BETWEEN THE CLIENT AND SERVER? They both
send replies and responses to the proxy, which forwards data to
the "other" end. When one party opens a data connection and sends
a PORT command to the proxy, the proxy allocates its own data
connection and sends its PORT command to the "other" end. The
proxy also copies data across the connections created in this way.
3.2 Characteristics of classical proxy configurations
Several IP internetworks may be linked using only classical proxy
technology. It is currently popular to link two specific IP
internetworks in this way: the Internet and some organization"s
"private" IP network. Such a proxy-based link is often the key
component of a firewall.
When this is done, several benefits and problems are introduced
for network administrators and users.
3.2.1 IP addressing and routing requirements.
The proxy system must be able to address all client and server
systems to which it may provide service. It must also know
valid IP routes to all these client and server systems.
Client and server systems must be able to address the proxy
system, and must know a valid IP route to the proxy system. If
the proxy system has several IP addresses (and often, several
physical network interfaces), the client and server systems
need only to be able to access ONE of the proxy system"s IP
addresses.
Note that client and server systems that use the proxy for
communication DO NOT NEED valid IP addressing or routing
information for systems that they reach through the proxy.
In this sense, it can be said that systems separated by a
classical proxy are isolated from each other in an IP
addressing sense and in an IP routing sense.
On the other hand, the classical proxy system (if running a
standard TCP/IP software stack) needs to have a single coherent
view of IP addressing and routing. If such a proxy system
interconnects two IP networks and two systems use the same IP
network/subnetwork number (one system on each network), the
proxy will only be able to address one of the systems.
This restriction can be removed by chaining classical proxies
(this is described later in section 3.2.6, "Interconnection of
conflicting IP networks").
Using a classical proxy for interconnection of IP
internetworks, it is also possible, with care, to achieve a
desirable "fail-safe" feature: no valid routing entries need to
exist for an internetwork which should be reached only through
the proxy (routing updates that could add such entries shout be
BLOCKED). If the proxy suddenly starts to behave like an IP
router, only one-way attacks become possible.
In other words, assume an attacker has control of the remote
internetwork and has found a way to cause the proxy to route IP
packets, or has found a way to physically bypass the proxy.
The attacker may inject packets, but the attacked internal
systems will be unable to reply to those packets. This
certainly does not make attacks infeasible (as exemplified by
certain holiday-period events in recent years), but it still
makes attacks more difficult.
3.2.2 IP address hiding
Application "sessions" that go through a classical proxy are
actually made of two complete sessions:
a) a session between the client and the proxy
b) a session between the proxy and the server
A device on the path sees only the client<->proxy traffic or
the proxy<->server traffic, depending where it is located. If
the two sessions actually pass through the same physical
network, a device on that network may see both traffics, but
may have difficulty establishing the relationship between the
two sessions (depending on the specific application and
activity level of the network).
A by-product of a classical proxy"s behavior is commonly known
as "address hiding". Equipments on some side of a classical
proxy cannot easily determine what are the IP addresses used on
another side of the proxy.
Address hiding is generally viewed as a Good Thing, since one
of the purposes of deploying proxies is to disclose as little
information about an internetwork as possible.
People who are in charge of gathering network statistics, and
who do not have access to the proxy system"s reports (if any)
may consider address hiding to be a Bad Thing, since the proxy
obscures the actual client/server relationships where the proxy
was inserted. All IP activity originates and terminates on the
proxy itself (or appears to do so).
In the same way, server software that accepts connections that
have gone through a classical proxy do not see the IP address
of the incoming client, unless this information is included in
the application protocol (and even if it is, in many cases, the
proxy will replace this information with its own address for
the protocol to be consistent). This makes server access
control unusable if it is based on client IP address checks.
3.2.3 DNS requirements
In most classical-proxy configurations, client systems pass the
desired server name (or address) to the proxy system WITHOUT
INTERPRETING IT. Because of this, the client system DOES NOT
REQUIRE to be able to resolve the name of the server system in
order to access it through a classical proxy. It only needs to
be able to resolve the name of the proxy (if referencing the
proxy system by name).
Because of this, it can be said that a classical proxy system
can offer DNS isolation. If two IP internetworks use completely
separate DNS trees (each with their own DNS root servers),
client software in one IP internetwork may still reference a
server name in the other IP internetwork by passing its name to
the classical proxy.
The classical proxy itself will not be able alone to resolve
DNS names in both environments (if running standard DNS
resolution software), since it will need to point to one or the
other of the two DNS "universes".
A well-known technique called "split-brain DNS" can be used to
relax this restriction somewhat, but such a technique
ultimately involves prioritizing one DNS environment over
another. If a DNS query can return a valid answer in both
environments, only one of the answers will be found by the
proxy.
3.2.4 Software requirements
A classical proxy application is a fairly simple piece of
software, often simpler than either a real client
implementation or a real server implementation. Such a program
may run on any system that supports normal TCP/IP connections,
and often does not require "system" or "superuser" privilege.
Classical proxy connections have no impact on normal server
software; the proxy looks like a normal client in most respects
except for its IP address and its "group" nature. All
connections from the network on the other side of the proxy
appear to come from the proxy, which poses problems if access
control by client system is desired.
Normal client software may access a classical proxy if the user
is willing or able to go through the extra steps necessary to
indicate the final server to the proxy (whatever they are).
Alternatively, modified (or newer) client software may be used
that knows how to negotiate transparently with the proxy.
3.2.5 Impact of a classical proxy on packet filtering
If packet filtering is needed around a classical proxy, the
packet filtering rules tend to be simplified, since the only
traffic needed and allowed will originate from or terminate on
the proxy (in an IP sense).
If the proxy starts behaving like an IP router, or if it is
physically bypassed, such filtering rules, if deployed
generally within an IP internetwork, will tend to prevent any
direct traffic flow between the "internal" internetwork and
"external" internetworks that are supposed to be only reachable
through the application proxy.
3.2.6 Interconnection of conflicting IP networks
By chaining classical proxies, it is possible to achieve some
interconnection of IP networks that have a high level of
conflict. In practice, this type of setup resolves IP
addressing conflicts much better than DNS conflicts. But DNS
conflicts are currently less of a problem because the DNS
"address space" is almost infinitely large (has anybody
calculated the possible DNS address space based on the RFC-
standard maximum host name length?).
Even though RFC1597 was never more than an informational RFC,
many organizations have been quietly following its suggestions,
for lack of an easier solution. Now assume two organizations
each use class A network number 10 on their network. Suddenly,
they need to interconnect. What can they do?
First possibility: one side changes network number (not as hard
as people think if properly planned, but this still represents
some work)
Second possibility: they merge the two numbers by renumbering
partially on each side to remove conflicts (actually harder to
do, but has the political advantage that both sides have to do
some work)
Third possibility: they communicate through chained classical
proxies:
+--------+ +--------+ +--------+ +--------+
/ Org. 1 Proxy Proxy / Org. 2
+ dmn1.com +---+ system +---+ system +---+ dmn2.com +
net 10 / 1 2 net 10 /
+--------+ +--------+ +--------+ +--------+
Both proxy 1 and 2 are standard systems running normal TCP/IP
software stacks. Their configuration is not typical, however:
a) The link between proxy 1 and proxy 2 may use any IP
network number that is not used (or not needed) on
either side. Nothing on Org.1 and Org.2"s networks
need to have an IP route to this network.
b) Proxy 1 has an IP route for network 10 that points to
Organization 1"s network, and does DNS resolution
(if required) using dmn1.com"s name servers.
c) Proxy 2 has an IP route for network 10 that points to
Organization 2"s network, and does DNS resolution
(if required) using dmn2.com"s name servers.
d) Proxy 1 and proxy 2 only require a host IP route to
each other for communication.
e) For this to be convenient, the classical proxy
applications must support the automatic selection of
a destination based on the client IP address.
f) On proxy system 1, the proxy software treats incoming
sessions from proxy system 2 in the normal way: the
"client" (proxy system 2) will be prompted in an
application-specific way for the final destination.
However, incoming sessions from Org.1 addresses are
immediately and automatically forwarded to proxy
system 2.
Proxy system 2 is configured similarly (that is,
connections coming from proxy 1 are prompted for a
target server name, connections from Org.2 addresses
are immediately and automatically forwarded to
proxy 1.
From a user"s point of view, the behavior of such a chained
proxy system is not very different from a single classical
application proxy:
a) A user on a client system with address 10.1.2.3
on Org.1"s network wishes to do an anonymous FTP to
"server.dmn2.com".
b) The user starts an FTP towards proxy 1. Proxy 1 sees
an incoming connection from an address in network 10,
so it immediately relays the connection to proxy 2.
c) Proxy 2 sees a connection coming from proxy 1, so it
prompts the client. The user sees the username prompt
and types (assuming FTP proxies that behave like TIS"s
ftp-gw):
anonymous@server.dmn2.com
This will be resolved IN THE CONTEXT OF Org. 2"S
NETWORK. The user can then complete the dialog and
use the FTP connection.
d) Note that this setup will work even if the client and
server have the EXACT SAME IP ADDRESS (10.1.2.3 in
our example).
If the proxy applications support selecting another
proxy based on the destination supplied by the client,
and if DNS domains are unique, more than two conflicting
IP networks can be linked in this way! Here is an
example configuration:
a) Four IP networks that all use network 10 are linked
by four proxy systems. The four proxy systems share a
common, private IP network number and physical link
(LAN or WAN).
b) A user on organization 1"s network wishes to access
a server on network 3. The user connects to its local
proxy (proxy 1) and supplies that target system name.
c) Proxy 1 determines, based on a configuration rule,
that the target system name is reachable by using
proxy 3. So it connects to proxy 3 and passes the
target system name.
d) Proxy 3 determines that the target system name is
local (to itself) and connects to it directly.
Security Implications of chained proxies
Obviously, when such "chained" configurations are built,
access control rules and logging based on a
final-client/final-server combination are difficult to
enforce, since the first proxy in the chain sees a
final-client/proxy relationship and the last proxy in
the chain sees a proxy/final-server relationship.
Doing better than this requires that the proxies be
capable of passing the "original-client" and
"final-destination" information back and forth in the
proxy chain for access control and/or logging purposes.
This requires the proxies to trust each other, and
requires the network path to be trusted (forging this
information becomes an Excellent attack).
Even if these problems were to be solved reliably, the
original goal of the proxy chains was to solve an IP
and possibly a DNS conflict. The "original-client" and
"final-destination" values may not have the same
meaning everywhere in the overall setup. Tagging the
information with a "universe-name" may help, assuming
it is possible to define unique universe names in the
first place. Obviously this topic requires more study.
4. Transparent application proxies
The most visible problem of classical application proxies is the need
for proxy-capable client programs and/or user education so that users
know how to use the proxies.
When somebody thought of modifying proxies in such a way that normal
user procedures and normal client applications would still be able to
take advantage of the proxies, the transparent proxy was born.
A transparent application proxy is often described as a system that
appears like a packet filter to clients, and like a classical proxy
to servers. Apart from this important concept, transparent and
classical proxies can do similar access control checks and can offer
an equivalent level of security/robustness/performance, at least as
far as the proxy itself is concerned.
The following information will make it clear that small organizations
that wish to use proxy technology for protection, that wish to rely
entirely on one proxy system for network perimeter security, that
want a minimal (or zero) impact on user procedures, and that do not
wish to bother with proxy-capable clients will tend to prefer
transparent proxy technology.
Organizations with one or more of the following characteristics may
prefer deploying classical proxy technology:
a) own a substantial internal IP router network, and wish to
avoid adding "external" routes on the network
b) wish to deploy "defence in depth", such as internal firewalls,
packet filtering on the internal network
c) wish to keep their DNS environment fully isolated from the
"other side" of their proxy system, or that fear that their
internal DNS servers may be vulnerable to data-driven attacks
d) use some IP networks that are in conflict with the "other side"
of their proxy system
e) wish to use proxy applications that are easily portable
to different operating system types and/or versions
f) wish to deploy multiple proxy systems interconnecting them
to the SAME remote network without introducing dynamic
routing for external routes on the internal network
4.1 Transparent proxy connection example
Let us go through an FTP sesssion again, through a "transparent"
proxy this time. We assume that the proxy system has two IP
addresses, two network interfaces, and two DNS names.
The proxy system is running a special program which knows how to
behave like an FTP client on one side, and like an FTP server on
the other side. This program is what people call the "proxy". This
program, being a transparent proxy, also has a very special
relationship with the TCP/IP implementation of the proxy system.
This relationship may be built in several ways, we will describe
only one such possible way.
We will assume that the proxy program is listening to incoming
requests on the standard FTP control port (21/tcp), although this
is not always the case in practice.
+---------------+ +----------+
/ IP
c.dmn1.com ----+ network(s) +----------+
(c1.c2.c3.c4) /
+---------------+ +----------+ +-----------------+
(p1.p2.p3.p4)
proxy1.dmn3.com

proxy2.dmn4.com
(p5.p6.p7.p8)
+---------------+ +----------+ +-----------------+
/ IP
s.dmn2.com ----+ network(s) +----------+
(s1.s2.s3.s4) /
+---------------+ +----------+
The user starts an instance of an FTP client program on the client
system "c.dmn1.com", and specifies a destination of "s.dmn2.com",
just like if it was reachable directly. On command-line systems,
the user typically types:
ftp s.dmn2.com
The client system needs to convert the server"s name to an IP
address (if the user directly specified the server by address,
this step is not needed).
Converting the server name to an IP address requires work to be
performed which ranges between two extremes:
a) the client system has this name in its hosts file, or has
local DNS caching capability and successfully retrieves the
name of the proxy system in its cache. No network activity
is performed to convert the name to an IP address.
b) the client system, in combination with DNS name servers,
generate DNS queries that eventually propagate close to the
root of the DNS tree and back down the server"s DNS branch.
Eventually, a DNS server which is authoritative for the
server system"s domain is queried and returns the IP
address associated with "s.dmn2.com" (depending on the
case, it may return this to the client system directly or
to an intermediate name server). Ultimately, the client
system obtains a valid IP address for s.dmn2.com.
+---------------+ +--------+
/ IP
c.dmn1.com --------+ network(s) +------------+
(c1.c2.c3.c4) /
+---------------+ +--------+ +-----------------+
A / (p1.p2.p3.p4)
address for / +-----+ proxy system
s.dmn2.com? / / (p5.p6.p7.p8)
/ / +-----------------+
/ /
/ / s.dmn2.com?
+--------+ / +--------+
+--------> DNS --+ +-------+ / IP
server / + network(s) +
+------------ X <---+ + /
s1.s2.s3.s4 +--------+ s1.s2.s3.s4 +--------+

+
+--------+
+ +-> DNS
server
+---- Y
+--------+
NOTE: In practice, DNS servers that are authoritative for
s.dmn2.com are highly likely to be located on the OTHER
side of the proxy system. This means that DNS queries
from the inside to the outside MUST be able to cross the
proxy system. If the proxy system wishes to provide
"address hiding", it must make these DNS queries
(originating from the inside) appear to come from the
proxy itself. This can be achieved by using a BIND-based
DNS server (which has some proxy capabilities) or some
simpler DNS proxy program. For full RFCcompliance,
the proxy system must be able to relay TCP-based queries
just like UDP-based queries, since some client systems
are rumored to ONLY use TCP for DNS queries.
The proxy system must be able to detect and block several
classes of attacks based on DNS which (if nothing else)
may cause denial of service:
a) attempts from the outside to return corrupt cache
entries to an internal DNS server
b) attempts to return DNS bindings which have no
relationship to the actual DNS query (some DNS
servers are vulnerable to this). The attacker"s goal
may be to prime the cache of internal DNS servers with
interesting entries, including entries for internal
DNS names that point to external IP addresses...
c) data-driven stuff similar in style to the "syslog
buffer overrun" type attacks.
Once the client system knows the IP address of the server system,
it attempts to establish a connection to the standard FTP
"control" TCP port on the server (port 21). For this to work, the
client system must have a valid route for the server"s IP address
THAT LEADS TO THE PROXY SYSTEM, and the proxy system must have a
valid route for the client"s IP address and the server"s IP
address. All intermediate devices that behave like IP gateways
must have valid routes for the client, the server, and usually the
proxy. If these devices perform packet filtering, they must ALL
allow the specific type of traffic required between C and S for
this specific application.
A
route to S

+-----------------+
+---------------+ (p5.p6.p7.p8)
c.dmn1.com proxy system
(c1.c2.c3.c4) (p1.p2.p3.p4)
+---------------+ +-----------------+

route to S route to C
V V

A A
route to C route to S

C S C
+----+ <-- +----+ --> +----+ <-- +----+
G1 -------- Gx -------- Gy --------- Gn
+----+ --> +----+ <-- +----+ --> +----+
S C S
At the start of the FTP session, a TCP packet with a source
address of C and a destination address of S travels to the proxy
system, expecting to cross it just like a normal IP gateway.
This is when the transparent proxy shows its magic:
The proxy"s TCP/IP software stack sees this incoming packets (and
subsequent ones) for a destination address that is NOT one of its
own addresses. Based on some criteria (a configuration file, for
example), it decides NOT to forward or drop the packet (which are
the only two choices an RFC-standard TCP/IP implementation would
have). The proxy system accepts the packet as if it was directed
to one of its own IP addresses.
In our example, the incoming packet is a TCP packet. Since
standard TCP/IP stacks store both a LOCAL and REMOTE IP address
field for each TCP connection, the transparent proxy may set the
LOCAL IP address field to the IP address that the client wants to
reach (s1.s2.s3.s4 in our example). The standard TCP/IP stack
probably needs to be modified to do this. UDP examples, although
not connection-based, could be handled in similar ways.
Once this is done, the actual FTP proxy application is invoked
since an incoming connection to TCP port 21 has occurred. It can
determine what is the final target destination instantly, since
the LOCAL IP address field of the connection contains the target
server"s IP address. There is no need for the proxy application
to ask the client what is the final target system.
Since the FTP proxy application knows the IP address of the target
system, it can choose to do two things:
a) Setup a session to the final target system, the more
frequent case.
b) Decide (based on some internal configuration data) that it
cannot reach the final target system directly, but must go
through a "classical" proxy. This seems technically
feasible, although no real transparent proxy system is
known to offer this capability. The actual value of such
a feature (if available) would need to be studied.
If the FTP proxy decides to connect directly to the target system,
it has the target system"s IP address. It may choose to do a
reverse lookup on the target IP address to obtain a target system
name (possibly needed for access control). If this process
involves DNS resolution, something like the following will happen:
+-----------------+
proxy1.dmn3.com
(p1.p2.p3.p4) +--------+
/ IP
proxy2.dmn4.com --------+ network(s) +------------+
(p5.p6.p7.p8) /
+-----------------+ +--------+ +---------------+
A / (s1.s2.s3.s4)
name for / s.dmn2.com
s1.s2.s3.s4? /
/ +---------------+
/
/
+--------+ s1.s2.s3.s4? +--------+
+--------> DNS ------------------> DNS
server server
+------------ X <------------------ Y
s.dmn2.com +--------+ s.dmn2.com +--------+
Once this is done and if the connection is allowed, the proxy
attempts to establish a connection to the standard FTP "control"
TCP port on the target server (port 21), using a technique
identical to a "classical" proxy. For this to work, the proxy
system must have a valid route to the server"s IP address, and the
server system must have a valid route to at least one of the
proxy"s IP address. All intermediate devices that behave like IP
gateways must have valid routes to both the proxy and the server.
If these devices perform packet filtering, they must ALL allow the
specific type of traffic required between the proxy and S for this
specific application.
+-----------------+
proxy1.dmn3.com
(p1.p2.p3.p4)
+----------------+
proxy2.dmn4.com s.dmn2.com
(p5.p6.p7.p8) (s1.s2.s3.s4)
+-----------------+ +----------------+

route to S route to P2
V V

A A
route to P2 route to S

P2 S P2
+----+ <-- +----+ --> +----+ <-- +----+
G1 -------- Gx -------- Gy --------- Gn
+----+ --> +----+ <-- +----+ --> +----+
S P2 S
The rest of the transparent proxy"s operation is very similar to
what would happen with a classical proxy.
4.2 Characteristics of transparent proxy configurations
Transparent proxy technology can be used to build the key
component of a "firewall", in a way quite similar to the way
classical proxy technology may be used. Several important details
of the architecture must be different, however.
4.2.1 IP addressing and routing requirements
The transparent proxy system must be able to address all client
and server systems to which it may provide service. It must
also know valid IP routes to all these client and server
systems.
Server systems must be able to address the proxy system, and
must know a valid IP route to the proxy system. If the proxy
system has several IP addresses (and often, several physical
network interfaces), the server systems need only to be able to
access ONE of the proxy system"s IP addresses.
Client systems MUST HAVE valid IP addressing and routing
information for systems that they reach through the proxy. For
example, in the common case where a transparent proxy is being
used to interconnect a private network and the Internet, the
private network will effectively need to use a default route
that points to the transparent proxy system. This is a specific
need of transparent proxy configurations.
Interconnecting two internetworks with multiple transparent
proxies (for load sharing or fail-over) can be accomplished by
using different techniques from what would be done for
classical proxies:
a) with multiple classical proxies to the same remote
network, clients can be configured to access different
proxies manually, or DNS-based techniques, such as
DNS load-balancing may be used to make clients
access a different proxy at different times.
b) with multiple transparent proxies to the same remote
network, the internal network must be able to provide
dynamic routing towards the proxies (routing updates
may need to be supplied by the proxies themselves).
Client systems (depending on topology) may not need
to see the route changes, but internal backbone
routers probably do.
It is clear that internetworks linked by a transparent proxy
cannot be fully isolated from each other in an IP addressing
and routing sense. The network on which client systems are
located must have effective valid routing entries to the remote
internetwork; these routing entries must point to the proxy.
The transparent proxy system (if running a vaguely standard
TCP/IP software stack) needs to have a single coherent view of
IP addressing and routing. If a proxy system interconnects two
IP networks and two systems use the same IP network/subnetwork
number (one system on each internetwork), the proxy will only
be able to address one of the systems. Even if the proxy is
able to manage multiple conflicting IP universes (if, for
example, one instance of a complete TCP/IP stack and its data
structures is bound to each of the proxy network interfaces),
the client systems will still have a problem: Why should it
send packets with this network number to the proxy since this
network number exists also on the internal internetwork?
Chaining transparent proxies does not seem at first glance to
solve IP conflicts like it does for classical proxies.
From a "security" fail-safe point of view, the transparent
proxy has an undesirable characteristic: the network being
protected must have valid routing entries to the remote
network(s). If the proxy fails (starts behaving like a non-
filtering IP router) or is physically bypassed, it is likely
that the internal network will be immediately able to reply to
"attacker" packets. The attacker does not need to modify
routing tables or to spoof internal IP addresses.
This is important for organizations that do not wish to place
ALL their confidence and protection into a proxy system (for
whatever reason).
4.2.2 IP address hiding
Application "sessions" that go through a transparent proxy are
actually made of two complete sessions:
a) a session between the client and the address of the
server, the session being "intercepted" by the proxy
b) a session between the proxy and the server
A device on the path sees either the client<->server traffic or
the proxy<->server traffic, depending where it is located. The
client<-"server" traffic is actually generated by the
transparent proxy. The two sessions SHOULD NEVER pass through
the same physical network, since in that case (due to the
routing requirements) a total bypass of the proxy at the IP
routing level may easily occur without being detectable.
Like classical proxies, transparent proxies accomplish a form
of IP address hiding. Client IP addresses are hidden from the
servers, since the servers see a session being initiated by the
proxy. Server IP addresses are NOT hidden from the clients
however, so that the illusion of transparency may be
maintained.
This difference implies that internal (client-side) network
statistics at the IP level will accurately reflect what outside
destinations are being accessed. This can be useful for
analyzing traffic patterns.
4.2.3 DNS requirements
In transparent proxy configurations, client systems MUST be
able to resolve server names belonging to remote networks. This
is critical since the proxy will determine the target server
from the destination IP address of the packets arriving from
the client. Because of this, the "client" internetwork needs to
have some form of DNS interconnection to the remote network. If
internal client and name server IP addresses must be hidden
from the outside, these DNS queries must also be proxied.
Of course, remote host name/address relationships may be stored
locally on the client systems, but it is well known that such
an approach does not scale...
Because of this, it can be said that a transparent proxy system
cannot offer DNS isolation. If two IP internetworks use
completely separate DNS trees (each with their own DNS root
servers), client software in one IP internetwork will not have
a way of finding name/address relationships in the "other" DNS
tree, and this information must be obtained in order to pass
the desired address to the transparent proxy.
The classical proxy itself (if running standard DNS resolution
software) will not be able alone to resolve DNS names in both
environments, since it will need to point to one or the other
of the two DNS "universes". Running multiple instances of DNS
resolution software can allow the proxy to do this, however.
Because of the requirement placed on some form of DNS
communication through the proxy, it is critical for the proxy
to be able to protect ITSELF, internal clients, and internal
name servers from data-driven attacks at the DNS level.
4.2.4 Software requirements
The big advantage of transparent proxies is that normal client
software may access remote servers with no modifications and no
changes to user procedures.
The transparent proxy application itself may not need to be
more complicated than a classical proxy application.
However, the proxy TCP/IP software stack cannot be a fully-
standard (well, today"s standard at least) TCP/IP stack, and
requires specific extensions:
a) the ability to specify ranges of IP addresses that
do not belong to the proxy itself, but for which
"intercept" processing will occur: if packets arrive
at the proxy with a destination IP address in those
ranges, the IP stack will not forward or drop the
packets; it will pass them up to application layers.
b) This mechanism requires that applications may obtain
both the IP address from which the packets come, and
the address to which the packets were going. Typical
IP stacks should already have the fields available
to store the info; it is a matter of updating them
properly for these "intercepted" packets.
c) In the case of "intercepted" TCP packets, the TCP
stack must support establishing TCP connections
where the "local" IP address is not one of the
proxy"s IP address.
Any TCP/IP software implementation should be modifiable to
perform these tasks. If a standard API becomes widely available
to drive these extensions, and if this API is generally
implemented, transparent proxies may become "portable"
applications.
Until this occurs, it must be assumed that implementors have
chosen different ways of accomplishing these functions, so that
today"s transparent proxy applications cannot be fully
portable. It also remains to be seen how much work is needed to
propagate these "extensions" to IPV6 software stacks.
4.2.5 Impact of a transparent proxy on packet filtering
The nature of a transparent proxy"s functionality makes it
difficult to deploy good packet filtering on the "inside" (or
client-side) of the proxy. The proxy will "masquerade" as all
the external systems. Because of this, internal packet filters
WILL TYPICALLY NEED TO ALLOW IP traffic between internal and
external IP addresses.
Depending on the actual security policy of the network, it may
be possible to do filtering based on protocol type and/or on
TCP bits (to filter based on connection setup direction), but
filtering that blocks external IP addresses CANNOT be deployed.
If the proxy starts behaving like an IP router, or if
physically bypassed, the practical limitations imposed on
internal packet filtering imply that a lot of direct traffic
between the inside and outside network will be allowed to flow.
Furthermore, as we have seen previously, the internal network
will have valid routing entries for external network numbers
that point to the proxy. If multiple proxies have been
deployed, the internal network may even HAVE TO TRUST routing
updates generated by the proxy.
In general, if an internal network wishes to communicate with
an external network through a transparent proxy, it MUST BE
FUNDAMENTALLY DESIGNED TO COMMUNICATE DIRECTLY with that
external network. This is true at the IP addressing level, at
the IP routing level, and at the DNS level. A proxy security
failure in this type of environment is likely to result in
immediate, total, and undetected accessibility of the internal
network by the external network.
4.2.6 Interconnection of conflicting IP networks
Unlike classical proxies, transparent proxies do not readily
seem useful in solving IP addressing conflicts.
If two internetworks use the same network number(s), systems
and routers in each internetwork will have valid routes to
these network numbers. If these routes are changed to point to
a transparent proxy, traffic that is meant to stay within the
same internetwork would start to flow towards the proxy. The
proxy will not be able to distinguish reliably between traffic
between systems of the same internetwork, and traffic which is
meant to cross the proxy.
A possible solution to this problem is described in section 6
of this document, "Improving transparent proxies".
5. Comparison chart of classical and transparent proxies
For those who do not like longish discussions of technical details,
here is a one-page summary of the strengths/weaknesses/differences of
classical and transparent proxies:
-----------------------------------------------------------------
Issue Classical Proxy Transparent Proxy
-------------------+---------------------+----------------------
IP addressing systems/gateways on systems/gateways on
each network need the "client" network
to address the proxy need to address the
remote networks

IP routing systems/gateways on systems/gateways on
each network need a the "client" network
valid routing entry also need routing
for the proxy entries for remote
entries

IP address hiding systems on each side systems on the
of the proxy are "client" side are
hidden from each hidden from the
other other sides

DNS full isolation resolution of outside
possible names by inside
systems is required

Proxy software runs on standard requires special
requirements TCP/IP stack; TCP/IP stack;
can be portable not 100% portable

Client software requires proxy- nothing more than for
requirements capable software a direct connection
or user education

User requirements must use proxy- nothing more than for
capable software or a direct connection
know how to use the
proxy

Packet filtering can filter out cannot filter out
"external" addresses "external" addresses

IP address can be done with no obvious way to
conflict chained proxies that get this to work
resolution support auto-connect
-----------------------------------