17th Apr 2002
COMMAND
AOL Instant Messenger direct connect feature leads to remote exploit
SYSTEMS AFFECTED
All versions of AOL Instant Messenger (up to 4.8 beta) on all platforms
of Windows.
PROBLEM
Noah Johnson posted this very fine analysis of AIM protocol weakness :
The problem arises in AIM's handling of embedded objects during direct
connections with other users. These 'direct connections' supposedly
make it easier for users to share multimedia with each other during
conversations. When a direct connection is to be made, the initiating
side acts as a server - on port 4443 - for the recipient's client to
connect to (if the request is accepted, of course). After this
connection is made, all activity between the two users is passed
through it, relieving the AIM server of its job for the time being.
When a user sends a picture or a sound to his buddy, an <IMG> tag is
appropriately inserted into the conversation source, while that file's
data lies in a separate <DATA> tag immediately proceeding the HTML
(see below). The client responds to this <IMG> tag either by
displaying the picture in the conversation or by displaying an icon of
the MIME/file-type that has been sent. Along with the standard
parameters of this <IMG> tag (HEIGHT, WIDTH, DATASIZE, ID, etc...),
the client also specifies the name and path of the original file that
was sent. This information is included in the "SRC" parameter, and is
the one of importance here. The client uses this information in a few
ways - including the default filename suggested when the user opts to
save whatever the hell was sent. So, why does anyone care?
One last nifty feature I forgot to mention... When the client parses
the file and recognizes it as a RIFF/WAVE type, it will play that file
instantly via the 'SndPlaySoundEx' API function. Instead of playing the
buffer directly from memory however, it is instead downloaded into the
Windows 'temp' directory and read from there. But for some odd reason,
THE ORIGINAL FILENAME IS USED FOR THIS TEMPORARY FILE! (Can you see
where this is heading??)
With very trivial parsing of this "SRC" parameter, AIM is left wide
open to the old ("..\..") directory traversal attack. So now we can
choose ANY path on our buddy's system to save our file to simply by
appropriately sculpting the "SRC" information. Here's an example of
what might be sent from the direct connection:
...
Data Headers (explained later)
<HTML><BODY>Hey, what's up?<IMG
SRC="\..\system\johnny.important_file" HEIGHT="0"
WIDTH="0" DATASIZE="50"
ID="1"></BODY></HTML><BINARY><DATA
ID=1">***WAVE FILE DATA
HERE***</DATA></BINARY>
...
Assuming that the temp directory on Johnny's computer is
[c:\windows\temp], this would write/overwrite whatever file was thus
specified. No obvious signs of this would be noted, as the WAVE icon
would never show up in Johnny's box (notice the HEIGHT and WIDTH
values).
Fooling the Client
==================
As noted earlier, AIM only saves this (semi-) temporary file when it
identifies it as valid RIFF/WAVE data and tries to play it. So what
good is this if we can only write WAVE files? ("HOLY SHIT! YOU MEAN WE
CAN OVERWRITE THEIR AIM SOUNDS?!") Unfortunately for Johnny, the client
only looks at the first 12 bytes before concluding it is indeed a valid
sound file. These bytes ideally look like this (see wave file
specifications):
Offset Data
0,1,2,3 ASCII 'RIFF'
4,5,6,7 DWORD Wave File Size ( <- can be
anything, we don't care;
neither does the client. )
8,9,A,B ASCII 'WAVE'
Keeping this as a base, we can then sculpt any file we want (sample
exploits next). As long as these 12 bytes are respected (actually, only
the ASCII parts of them really matter), the file is saved.
The fact that the header is already determined severely limits the
danger of this bug (we cannot, for example, create an executable file).
There are a few ways around this...
Exploiting
==========
Besides the potential of overwriting files, there are plenty of file
types that can be bullshitted in the first 12 bytes without losing
functionality of the entire thing (think scripts, for example). By
placing these files in the Start-Up folder, you can see where this
could lead...
For example:
A sample .BAT file
C:\Windows\Start Menu\StartUp\exploit.bat
{
RIFF----WAVE----
Any dos command here
Any dos command here
etc...
exit
}
While I originally assumed the '-'s would be the ASCII character 0x08
(backspace, which would delete the 'WAVE' and 'RIFF' so the shell would
not try to execute these commands, give an error and then quit), it
turns out that even if an error is encountered ('Invalid command or
file name'), it will not stop the rest of the script from executing
upon start-up.
Another example:
A sample .VBS file
C:\Windows\Start Menu\StartUp\exploit.vbs
{
RIFF____WAVE = ""
Any vbs code here
Any vbs code here
etc...
}
Here, the first line (RIFF____WAVE = "") is a valid declaration, which
means no error will occur in scripting and anything else can be put in
afterwards.
A Few Things to Note
====================
Because AIM only allows multimedia to be sent in Direct Connection
sessions (not quite true, see note on Buddy Icons), and because AIM
filters out HTML tags (such as our <IMG> tag here) when inserted
directly into an IM message, in order to exploit this bug a user must
be able to control the raw data being sent. The most obvious way is to
simply take control of the port that the connection is to be made on
(before the client does) and sculpt the data on your own...
But sequence numbers are used in the transmission of data during these
sessions. The base sequence number is randomly selected by the
initiating side, and is sent to the buddy's client as part of the
request for a connection (Port # is sent at the same time). This
sequence is incremented every time an entire chunk of data is sent
(which could span any number of packets; the size of every data piece
is sent right after the sequence number). If the sequence number or
data size is wrong, the direct connection will be closed immediately.
You can experiment on your own on port 4443.
The simplest way I can think of to test this bug (avoiding dealing with
sequences and protocols) is to let the client handle everything - the
connections and all - and manually edit the conversation source [with
the 'wave file' already inserted] just before it's sent out
(SendWindowMessage (WM_SETTEXT) on the RichTextBox or use SoftIce and
search for the data in memory or something to that nature). I won't
publish a full-out exploit; I think you get the picture.
A Note on Buddy Icons
=====================
If you send this same type of forged WAVE file as a Buddy Icon (no
direct connection needed), the client does the same thing described
earlier, even before the "Accept Buddy Icon" prompt appears. I don't
know if this is intentional on the part of AOL, but as far as I can
tell it's impossible to specify your own file name. Instead the client
uses some default name of its own. It's just odd, because I was under
the impression that Buddy Icons were supposed to be graphic files >>>
(Possible DoS in this - send a corrupt wave file to buddy's computer? I
don't know, anyone care to investigate?)
Also, when a buddy is sent your Buddy Icon, the file is downloaded onto
their computer and saved WHETHER THEY CHOSE TO ACCEPT IT OR NOT.
Located in their AIM user's 'picture' folder and named 'T.id',
THEORETICALLY one could use this place to store an executable file to
be renamed and executed later via script. Just a thought, but you
didn't hear that from me...
SOLUTION
Never accept a direct connection from anyone you don't trust.
No fix yet.