When VLAD started working on the Win95 problem (back when WinSurfer was
  released), all we had to work with was the win32s update to win3.1 that
  allowed win3.1 users to run (not very successfully) 32 bit win95/NT
  applications.  Tracing through the code one day we noticed that every
  exe started with loading the PE header with eax pointing to it and
  jumping to the entrypoint (jmp [eax + 28h]).  So we changed the
  entrypoint RVA to point to our code, pushed eax, did our code, popped
  eax, and jumped to [eax + 28].  It worked, we had a win32s infector,
  then we noticed we didn't actually have any way to infect.  There were
  no ints and we really felt out of our depth and then to top it all off
  we came across the ultimate way to go resident without infecting the
  win3.1 shell and so PH33R was born.  About this time I left the scene
  in an unexplainable burst of insane "oh my god I'm going to fail uni"
  panic and that was about the end of the win32s/95/NT project.

  Just recently, uni ended.  I was on holidays, bored out of my skull,
  and running win95, so I decided I'd give it another go.  To my surprise
  the jmp [eax + 28] was gone, it was just a byproduct of win32s.  Instead
  I just calculated the distance between the end of the exe and the
  entrypoint, calculated the start of the virus code and subtracted the
  two.  It worked, I had a win95/NT infector.  That is, 'cept for a little
  matter of calling API...


  The Day the World Changed
  -------------------------

  In the beginning there were memory locations and we manipulated what
  our computer did by peeking and poking.  And it was good.  But this
  was not to last, for on the horizon was the "port" and slowly we
  learnt to communicate with attached devices by puting values in, and
  pulling values out, of specific ports.

  After a while the computer was no longer young and we needed
  something to keep things in check.  We needed a way to communicate
  indirectly with the hardware and thus the interrupt was born.  Not
  everyone liked the interrupt system but soon we all learnt to live
  with it, to manipulate it, and use it to our advantage.

  Everyone that is, except one group at Microsoft - the guys who were
  writing a "revolutionary" GUI called "Windows".  These guys were
  looking for a way to split all the functions that were once provided
  by interrupts into seperate, shared files called Dynamic Link
  Librarys (Dll's).  And so the Application Program Interface (API) was
  born.  API's made calling functions just that little bit harder and
  admittedly sometimes completely impossible.  But the guys at Microsoft
  had not yet taken away our interrupts and we still had (some) control
  over the system with DPMI.

  Then the guys at Microsoft did it again.  Their latest concoction
  eliminated interrupts and ports and yes, even memory locations.  For
  "Windows 95/NT" is a true non-preemptive multitasking system or in
  other words - a bully.  This ogre pushes programs around, squeezes
  them into confined spaces, locks them out of restricted areas and,
  worst of all, forces them into using a new, impossibly complex, method
  of calling API.

  Which brings us to...


  One drunk night at Microsoft
  ----------------------------

  When the guys at Microsoft get drunk, they truly get drunk.  But rather
  than running around with a cop's panties on their collective head and
  stealing traffic cones, they designed a new call construct.

  In the spirit of Windows 3.1, the guys at Microsoft chose against using
  interrupts.  In some ways this was good - programmers dont need to check
  to see if dll's are in memory (although they could), and they can state
  what dll's they need.  The way that Microsoft achieved this is through
  the Import and Export Tables.  The Import table lets the programmer state
  what functions they need from which dll's.  The Export table lets the
  provider of the dll specify what functions it provides.

  The problem is, once the import table has been written it is set in stone.
  There is no space to add any new entries and you cant move anything around.
  Why ?  Well, mainly because of a certain jump table that can be ANYWHERE
  in the exe.  To call an API the code calls a hardcoded address.  This
  address points to an indirect jump which points to an entry in the import
  table that is filled in (when the executable is loaded) with an address to
  a similar jump table in the dll that pushes a value and jumps to the
  function dispatcher which uses the values that are inserted in the dll's
  export table.  To add an entry to the import table you would have to
  move the current entries around so that you could fit it in.  This means
  you have to change the jump table to point to the relocated entries, which
  is impossible as the jump table can be ANYWHERE in the file and cannot be
  located.

  The New Frontier
  ----------------

  Windows 95 is a whole new platform.  It's a new challenge and although
  a lot of things have been stacked against you - it IS possible.  So
  get in there and start researching.  Research will be the end-all of
  the VX/AV war, one way or the other.

  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

  Function              June            August
  --------------------------------------------------
  GetCurrentDir       BFF77744         BFF77744
  SetCurrentDir       BFF7771D         BFF7771D
  GetTime             BFF9D0B6         BFF9D14E
  MessageBox          BFF638D9         BFF638D9
  FindFile            BFF77893         BFF77893
  FindNext            BFF778CB         BFF778CB
  CreateFile          BFF77817         BFF77817
  SetFilePointer      BFF76FA0         BFF76FA0
  ReadFile            BFF75806         BFF75806
  WriteFile           BFF7580D         BFF7580D
  CloseFile           BFF7BC72         BFF7BC72


  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:

          push offset lpszProc
          push dword ptr [hModule]
          call GetProcAddress

  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.

   RVA             = ((prev RVA + prev Virtual Size)/OBJ Alignment+1)
                                                               *OBJ Alignment
   Virtual Size    = ((size of virus+buffer any space)/OBJ Alignment+1)
                                                               *OBJ Alignment
   Physical Size   = (size of virus/File Alignment+1)*File Alignment
   Physical Offset = prev Physical Offset + prev Physical Size
   Object Flags    = db 40h,0,0,c0h
   Entrypoint RVA  = RVA

   Increase the 'number of objects' field by one.

   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.


  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.