Infection of Portable Executables
by
Qark and Quantum [VLAD]
The portable executable format is used by Win32, Windows NT and Win95,
which makes it very popular and likely to become the dominant form of
executable sometime in the future. The NE header used by Windows 3.11
is completely different to the PE header and the two should not be
confused.
None of the techniques in this document have been tested on Windows NT
because no virus writer (we know) has access to it.
At the bottom of this document is a copy of the PE format, which is not
easy to follow but is the only reference publicly available. Turbo
Debugger 32 (TD32) is the debugger used during the research of this
text, but SoftIce'95 also does the job.
Calling Windows 95 API
����������������������
A legitimate application calls win95 api by the use of an import
table. The name of every API that the application wants to call is
put in the import table. When the application is loaded, the data
needed to call the API is filled into the import table. As was
explained in the win95 introduction (go read it), we cannot modify
this table due to Microsoft's foresight.
The simple solution to this problem is to call the kernel directly.
We must completely bypass the Win95 calling stucture and go straight
for the dll entrypoint.
To get the handle of a dll/exe (called a module) we can use the API
call GetModuleHandle and there are other functions to get the
entrypoint of a module - including a function to get the address of an
API, GetProcAddress.
But this raises a chicken and egg question. How do I call an API so I
can call API's, if I can't call API's ? The solution is to call api
that we know are in memory - API that are in KERNEL32.DLL - by calling
the address that they are always located at.
Some Code
���������
A call to an API in a legitimate application looks like:
call APIFUNCTIONNAME
eg. call CreateFileA
This call gets assembled to:
db 9ah ; call
dd ???? ; offset into jump table
The code at the jump table looks like:
jmp far [offset into import table]
The offset into the import table is filled with the address of the
function dispatcher for that API function. This address is obtainable
with the GetProcAddress API. The function dispatcher looks like:
push function value
call Module Entrypoint
There are API functions to get the entrypoint for any named module but
there is no system available to get the value of the function. If we
are calling KERNEL32.DLL functions (of which are all the functions
needed to infect executables) then we need look no further than this
call. We simply push the function value and call the module
entrypoint.
Snags
�����
In the final stages of Bizatch we beta tested it on many systems.
After a long run of testing we found that the KERNEL32 module was
static in memory - exactly as we had predicted - but it was at a
different location from the "June Test Release" to the "Full August
Release" so we needed to test for this. What's more, one function
(the function used to get the current date/time) had a different
function number on the June release than it did on the August release.
To compensate I added code that checks to see if the kernel is at one
of the 2 possible locations, if the kernel isn't found then the virus
doesn't execute and control is returned to the host.
Addresses and Function Numbers
������������������������������
For the June Test Release the kernel is found at 0BFF93B95h
and for the August Release the kernel is found at 0BFF93C1Dh
Using a debugger like Turbo Debugger 32bit found in Tasm 4.0, other
function values can be found.
Calling Conventions
�������������������
Windows 95 was written in C++ and Assembler, mainly C++. And although
C calling conventions are just as easy to implement, Microsoft didn't
use them. All API under Win95 are called using the Pascal Calling
Convention. For example, an API as listed in Visual C++ help files:
FARPROC GetProcAddress(
HMODULE hModule, // handle to DLL module
LPCSTR lpszProc // name of function
);
At first it would be thought that all you would need to do is push the
handle followed by a pointer to the name of the function and call the
API - but no. Due to Pascal Calling Convention, the parameters need
to be pushed in reverse order:
Using a debugger like Turbo Debugger 32bit we can trace the call (one
step) and follow it to the kernel call as stated above. This will
allow us to get the function number and we can do away with the need
for an entry in the import table.
Infection of the PE Format
��������������������������
Finding the beginning of the actual PE header is the same as for NE
files, by checking the DOS relocations for 40h or more, and seeking to
the dword pointed to by 3ch. If the header begins with a 'NE' it is a
Windows 3.11 executable and a 'PE' indicates a Win32/WinNT/Win95 exe.
Within the PE header is 'the object table', which is the most important
feature of the format with regards to virus programming. To append code
to the host and redirect initial execution to the virus it is necessary to
add another entry to the 'object table'. Luckily, Microsoft is obsessed
with rounding everything off to a 32bit boundary, so there will be room
for an extra entry in the empty space most of the time, which means it
isn't necessary to shift any of the tables around.
A basic overview of the PE infection:
Locate the offset into the file of the PE header
Read a sufficient amount of the PE header to calculate the full size
Read in the whole PE header and object table
Add a new object to the object table
Point the "Entry Point RVA" to the new object
Append virus to the executable at the calculated physical offset
Write the PE header back to the file
To find the object table:
The 'Header Size' variable (not to be confused with the 'NT headersize')
is the size of the DOS header, PE header and object table, combined.
To read in the object table, read in from the start of the file for
headersize bytes.
The object table immediately follows the NT Header. The 'NTheadersize'
value, indicates how many bytes follow the 'flags' field. So to work
out the object table offset, get the NTheaderSize and add the offset
of the flags field (24).
Adding an object:
Get the 'number of objects' and multiply it by 5*8 (the size of an object
table entry). This will produce the offset of the space within which
the new virus object table entry can be placed.
The data for the virus' object table entry needs to be calculated using
information in the previous (host) entry.
Write the virus code to the 'physical offset' that was calculated, for
'physical size' bytes.
Notes
�����
Microsoft no longer includes the PE header information in their developers
CDROMs. It is thought that this might be to make the creation of
viruses for Win95 less likely. The information contained in the next
article was obtained from a Beta of the Win32 SDK CDROM.
Tools
�����
There are many good books available that supply low level Windows 95
information. "Unauthorized Windows 95", although not a particularly
useful book (it speaks more of DOS/Windows interaction), supplies
utilities on disk and on their WWW site that are far superior to the
ones that we wrote to research Win95 infection.
