|
COMMAND Preventing buffer exploits discussion SYSTEMS AFFECTED Windows platforms PROBLEM - see below - SOLUTION David Litchfield [david@ngssoftware.com] suggests : http://www.ngssoftware.com/ Defeating Exploits ****************** The ideas in this "paper" present a method for defeating exploits; not the actual vulnerability. Before getting to the details let's consider slammer (again). What made slammer so successful? The overriding factor that made slammer so successful was it's ability to spread. What made it's spread a foregone conclusion was the fact that every vulnerable SQL Server/MSDE had a "jmp esp" instruction at address 0x42B0C9DC. This was the address that was used to gain control of the SQL Server's path of execution to a point where the worm's payload, the "arbitrary code", would be executed. This address is in a dynamic link library (DLL) , sqlsort.dll which has an "image base" of 0x42AE0000. Every image file, DLL or executable, has an "Image Base" and this base is the preferred location where the file should be loaded into memory by the Windows Loader. [I don't want to digress, here, as to what happens if there's a conflict. See the references at the end.] Now if this Image Base on one particular system had been 0x42AF0000 then the worm would have failed to infect this particular box; the "jmp esp" instruction that should've been at 0x42B0C9DC on this system would be found at 0x42B1C9DC so the worm would have been off target. The SQL Server running on this system, whilst still being "vulnerable" to the buffer overflow vulnerability would have been invulnerable to this worm. Sure - the SQL Server may have crashed - but it would not have been compromised. It's like sickle cell. Someone born with the gene that causes sickle cell anaemia, a blood disorder that affects many people of a West African origin, or carriers of the gene, sickle cell trait, do not suffer from the ill affects of malaria, a disease caused by a parasite and most commonly spread by mosquitoes. Whilst someone with sickle cell trait can still catch malaria, the gene mutates the haemoglobin in their blood in such a way that they are invulnerable to the debilitating side effects and syptoms of the disease such as mental confusion, coma and death. There is an obvious evolutionary advantage to sickle cell trait; remember that the evolution of the species cares not about how long a person lives, only that they live long enough to pass on their genes. (Those with anaemia may suffer from crises, periods of acute pain so the trade off is somewhat questionable.) In areas where malaria is a common cause of death, being a carrier of the sick cell gene can help ensure that this person lives long enough to have progeny. This is Darwinian natural selection in progress. Rebasing ******** The problem with operating systems is that they all have pretty much the same "genetic code" which makes each and every one of them vulnerable to a new exploit. So we need to make them different and this can be achieved through rebasing. Rebasing is the process of changing the Image Base of an image file. By doing this the DLL/EXE is loaded into a different location in the virtual address space. Going back to Slammer, had I have rebased sqlsort.dll giving it a new base of 0x41410000 my box would have been invulnerable to the worm. If another worm were written, though, that used an address that contained a "jmp esp" instruction in kernel32.dll then I would be vulnerable. So I rebase kernel32.dll. But then another worm uses another DLL so I rebase that one, too. Eventually I've rebased all of the DLLs used by SQL Server mutating it's "genetic code", making it considerably different to any other SQL Server install on the planet. In fact if I rebase every DLL on my system and every executable then I can make my box almost invulnerable to a given exploit, past, present or future. It's not that my box is invulnerable to a buffer overflow vulnerability - it's just invulnerable to the exploits for it. To gain control of a system protected in such a way would require that the author of the exploit know the location of loaded DLLs. So how easy is it to rebase DLLs and executables? Very. Microsoft have provided a function to do this, ReBaseImage(), exported by imagehlp.dll. If you rebase an image the new base must be on a 64K boundary - i.e. if the image base mod 64000 !=0 the base is not valid. The only other problem is Windows File Protection. Once you've rebased a copy of the DLL you need to copy the new DLL over the old one but Windows File Protection won't allow you to do this. To get around the problem use the MoveFileEx and specifying the MOVEFILE_DELAY_UNTIL_REBOOT flag. Doing this will add a registry value, "PendingFileRenameOperations" to HKLM\System\CurrentControlSet\Control\Session Manager\. You then need to add another DWORD value "AllowProtectedRenames" and set it to 1. Then restart the system. On reboot the new DLLs, with their new image bases, will be loaded. For example - here is sample output of listdlls after kernel32.dll and ws2_32.dll have been rebased. Copyright (C) 1997-2000 Mark Russinovich http://www.sysinternals.com ---------------------------------------------------------------------------- -- WINLOGON.EXE pid: 208 Command line: winlogon.exe Base Size Version Path 0x01000000 0x2e000 \??\C:\WINNT\system32\winlogon.exe 0x77f80000 0x7b000 5.00.2195.5400 C:\WINNT\System32\ntdll.dll 0x78000000 0x46000 6.01.9359.0000 C:\WINNT\system32\MSVCRT.dll 0x4a4a0000 0xb1000 5.00.2195.6011 C:\WINNT\system32\KERNEL32.dll 0x77db0000 0x5b000 5.00.2195.5992 C:\WINNT\system32\ADVAPI32.dll 0x77d30000 0x71000 5.00.2195.5419 C:\WINNT\system32\RPCRT4.dll 0x54530000 0x13000 5.00.2195.4874 C:\WINNT\system32\WS2_32.dll .. .. Now all the way through this I've been saying things like "almost invulnerable" etc. Here's the reason. For some vulnerabilities it may be sufficient to overwrite a saved return address, function pointer or whatever by only a few bytes. For example assume a saved return address is 0x44784500 and at address 0x44784536 is a "jmp ebx" instruction and ebx points to our code. Then we only need to overwrite the saved return address by 1 byte - with 0x36. So knowledge of the DLL load address is not needed. However, this scenario is going to happen so infrequently (if ever) that it does not detract from the idea of rebasing your system. There may other ways to bypass this method. Some ideas to further help prevent exploits from working. Use addresses such as 0x**000000 or 0x00**0000 for the new image base. With there being a NULL in much of the image's address space this will help. (This of course won't make a difference with unicode overflows) Ensure at least one (core) DLL has a base of 0x00119400 . This will ensure that a common stack location 0x00120000 has been assigned forcing the OS to chose another location for the stack. You get the idea. MSDN Info ReBaseImage() http://msdn.microsoft.com/library/default.asp?url=/library/en-us/debug/base/rebaseimage.asp MoveFileEx() http://msdn.microsoft.com/library/default.asp?url=/library/en-us/fileio/base/movefileex.asp Update (05 Februrary 2003) ====== Thomas [dullien@gmx.de] comments : --snip-- Rebasing everything is something you're not very likely to achieve. Hardly any commercial software has executables which still contain valid relocation information -- which means that you can rebase all DLL's as much as you want, the main EXE (which is always mapped at 0x00400000 and cannot be remapped) will be present & can be used for exploitation. Unless you rebase the complete address space you remain vulnerable. Furthermore, rebasing might not be sufficient, as there's less than 32k different bases -- if the service restarts cleanly brute force is definitely an option. So you need full randomization. Heap corruptions allow an attacker to write arbitrary data to arbitrary locations -- so he can patch his own "jmp ebx" or whatever to whereever he wishes. Unless you implement something PaX-like for writable/executable pages, you're still vulnerable. And the majority of all buffer overruns _are_ heap corruptions. Oh, and there's always the static mapping of the TEB's under Windows. So the solution you're proposing a) Will only work against a small subset of all closed-source-applications (those with relocatable main .exe) b) Will even then only protect you against vanilla stack smashes, and offer 0 protection against heap corruptions or format string bugs c) Will be suspectible to brute-force attacks on your address space (which cannot be more complex than 2^15 ... hardly a "hard" task) -Also- Someone points out : There is a tool called "ReBase" shipped with Visual C++ and Visual C++.NET. http://msdn.microsoft.com/library/default.asp?url=/library/en-us/tools/perfutil_2z39.asp " Rebase is a command-line tool that you can use to specify the base addresses for the DLLs that your application uses. " " Alternatively, you can use the ReBaseImage function. " Update (06 February 2003) ====== Riley Hassell [rhassell@eeye.com], Security Research Associate of eEye Digital Security argues : So the course of this talk with most likely go into generating a totally dynamic address space and once again, end in another theoretical solution, to an overly complex problem. Defeating Rebasing ------------------------------------- Many operating systems with fault handling features and refined multitasking, reference address spaces with segments to permit these features and aid general performance. The majority of this behavior can be studied by following process creation and task switching. Start from the user API and step through until the entry point of the executable is reached. TEB: Thread Environment Block TIB: Thread Information Block PEB: Process Environment Block The TEB/TIB fs:[] segment references originated in the OS2 days and have since passed down into 9x client systems, and of course, Windows NT. During the process creation the PEB and TIB are initialized into the new virt and can be referenced by the fs segment. Modifying the address space referenced by fs and how fs is setup is possible... but by the time you're done you're designing a new operating system. As Ryan might say... It's all data ;) The data referenced in this delta can be referenced during a ret. You just need to find a set of bytes that forms the needed instruction. You may be able to modify this arena in a way that you can insert your own instructions. Maybe some of the TLS storage can be controlled by supplying malformed sizes in your exploit session... ;) Note: PEB locking pointers can be overwritten with format bugs and control structure based heap overflows. 7FFDF000 00010000 7FFDF004 FFFFFFFF 7FFDF008 01000000 <- Executable image base ;) 7FFDF00C 00071E90 7FFDF010 00020000 7FFDF014 00000000 7FFDF018 00070000 7FFDF01C 77FCF170 <- PEB fast Lock entry point 7FFDF020 77F8313C <- PEB lock entry point 7FFDF024 77F8316D <- PEB unlock entry point Brett Moore wrote me several months ago with a very interesting exploit concept using multiple writes. The first write you insert your needed instruction into writeable memory somewhere. The second write you overwrite some writeable entry point or hook, with the address of the inserted instruction. There's not much out there if you're interested in learning about the TIB and the PEB. If you really want to understand these structures and general loading behavior, learn Polish and Russian, then hit up some VX'er archives. If you end up talking to any of them, tell them that somebody is trying to stop exploits by rebasing dll's :) Rebasing... There's a reason why relocation sections exist. While doing your own relocations is possible, the design of such a system is extremely, and I'll say again, extremely complex. Just differentiating all the instructions from data is a fairly painful process. Maybe the ETCH guys did this at one point but as far as I know this has been a big hurdle in image modifcation for quite some time. Michal Zalewski provided some great examples of issue's you'll run into: MZ> Also, what if I wanted to pass a value 4325404 (0x42001c) to this MZ> function, and it is not a pointer, only looks this way? For example,some MZ> FOO_ASYNC flag is defined as 0x400000, FOO_LOCK as 0x020000, and voila,OR MZ> them and you have "a pointer". MZ> In other cases, say, with register calls, it is getting even nastier, MZ> because even if, one way or another, you managed to find out how every MZ> single function is going to use its parameters (not likely), register MZ> calls are still black magic. GetProcAddress ACL's... > > It is possible to intercept every call to GetProcAddress and determine > whether or not the call should be authorized based on a predetermined list > of known valid callers (runtime call stack analysis). Simply rewrite a micro GetProcAddress. GetProcAddress is basically an overstuffed RVA engine. This is defeated by "The payload brings the tools concept". Most userland hooking schemes can also easily be bypassed by using direct gates "Ex: Interupts Gates". You could also intercept a thread that has the neccesary privelege by snagging a hook in it's path. > This list of authorized callers must be constructed through the use of forensic profiling > tools in the case of other people's binaries, but can be constructed with > the help of additional API calls in the case of one's own code. Call a > profiling/tracing API before calling GetProcAddress. After compilation but > before deployment to production boxes you simply execute the code in profile > mode to generate a list of authorized callers. This list is then configured > as a static security setting adhered to by the security layer that sits > between GetProcAddress and the rest of the virtual world. Who's an authorized caller? Someone who has a "safe" caller address on the stack.... If a the attacker start's offering instructions to your CPU... kiss your ass goodbye. Research AV/VX trends from the late 80's and early 90's. Update (08 Februrary 2003) ====== Alex Fedotov proposed the following PEB/TEB data structures : http://groups.google.com/groups?hl=en&lr=&ie=UTF-8&oe=UTF-8&safe=off&q=_NT_TEB&btnG=Google+Search typedef struct _CURDIR { UNICODE_STRING DosPath; PVOID Handle; } CURDIR, *PCURDIR; typedef struct RTL_DRIVE_LETTER_CURDIR { USHORT Flags; USHORT Length; ULONG TimeStamp; UNICODE_STRING DosPath; } RTL_DRIVE_LETTER_CURDIR, *PRTL_DRIVE_LETTER_CURDIR; typedef struct _PEB_FREE_BLOCK { struct _PEB_FREE_BLOCK* Next; ULONG Size; } PEB_FREE_BLOCK, *PPEB_FREE_BLOCK; /* RTL_USER_PROCESS_PARAMETERS.Flags */ #define PPF_NORMALIZED (1) typedef struct _RTL_USER_PROCESS_PARAMETERS { ULONG MaximumLength; // 00h ULONG Length; // 04h ULONG Flags; // 08h ULONG DebugFlags; // 0Ch PVOID ConsoleHandle; // 10h ULONG ConsoleFlags; // 14h HANDLE InputHandle; // 18h HANDLE OutputHandle; // 1Ch HANDLE ErrorHandle; // 20h CURDIR CurrentDirectory; // 24h UNICODE_STRING DllPath; // 30h UNICODE_STRING ImagePathName; // 38h UNICODE_STRING CommandLine; // 40h PWSTR Environment; // 48h ULONG StartingX; // 4Ch ULONG StartingY; // 50h ULONG CountX; // 54h ULONG CountY; // 58h ULONG CountCharsX; // 5Ch ULONG CountCharsY; // 60h ULONG FillAttribute; // 64h ULONG WindowFlags; // 68h ULONG ShowWindowFlags; // 6Ch UNICODE_STRING WindowTitle; // 70h UNICODE_STRING DesktopInfo; // 78h UNICODE_STRING ShellInfo; // 80h UNICODE_STRING RuntimeInfo; // 88h RTL_DRIVE_LETTER_CURDIR DLCurrentDirectory[0x20]; // 90h } RTL_USER_PROCESS_PARAMETERS, *PRTL_USER_PROCESS_PARAMETERS; #define PEB_BASE (0x7FFDF000) typedef struct _PEB_LDR_DATA { ULONG Length; BOOLEAN Initialized; PVOID SsHandle; LIST_ENTRY InLoadOrderModuleList; LIST_ENTRY InMemoryOrderModuleList; LIST_ENTRY InInitializationOrderModuleList; } PEB_LDR_DATA, *PPEB_LDR_DATA; typedef VOID STDCALL (*PPEBLOCKROUTINE)(PVOID); typedef struct _PEB { UCHAR InheritedAddressSpace; // 00h UCHAR ReadImageFileExecOptions; // 01h UCHAR BeingDebugged; // 02h UCHAR Spare; // 03h PVOID Mutant; // 04h PVOID ImageBaseAddress; // 08h PPEB_LDR_DATA Ldr; // 0Ch PRTL_USER_PROCESS_PARAMETERS ProcessParameters; // 10h PVOID SubSystemData; // 14h PVOID ProcessHeap; // 18h PVOID FastPebLock; // 1Ch PPEBLOCKROUTINE FastPebLockRoutine; // 20h PPEBLOCKROUTINE FastPebUnlockRoutine; // 24h ULONG EnvironmentUpdateCount; // 28h PVOID* KernelCallbackTable; // 2Ch PVOID EventLogSection; // 30h PVOID EventLog; // 34h PPEB_FREE_BLOCK FreeList; // 38h ULONG TlsExpansionCounter; // 3Ch PVOID TlsBitmap; // 40h ULONG TlsBitmapBits[0x2]; // 44h PVOID ReadOnlySharedMemoryBase; // 4Ch PVOID ReadOnlySharedMemoryHeap; // 50h PVOID* ReadOnlyStaticServerData; // 54h PVOID AnsiCodePageData; // 58h PVOID OemCodePageData; // 5Ch PVOID UnicodeCaseTableData; // 60h ULONG NumberOfProcessors; // 64h ULONG NtGlobalFlag; // 68h UCHAR Spare2[0x4]; // 6Ch LARGE_INTEGER CriticalSectionTimeout; // 70h ULONG HeapSegmentReserve; // 78h ULONG HeapSegmentCommit; // 7Ch ULONG HeapDeCommitTotalFreeThreshold; // 80h ULONG HeapDeCommitFreeBlockThreshold; // 84h ULONG NumberOfHeaps; // 88h ULONG MaximumNumberOfHeaps; // 8Ch PVOID** ProcessHeaps; // 90h PVOID GdiSharedHandleTable; // 94h PVOID ProcessStarterHelper; // 98h PVOID GdiDCAttributeList; // 9Ch PVOID LoaderLock; // A0h ULONG OSMajorVersion; // A4h ULONG OSMinorVersion; // A8h ULONG OSBuildNumber; // ACh ULONG OSPlatformId; // B0h ULONG ImageSubSystem; // B4h ULONG ImageSubSystemMajorVersion; // B8h ULONG ImageSubSystemMinorVersion; // C0h ULONG GdiHandleBuffer[0x22]; // C4h PVOID ProcessWindowStation; // ??? } PEB, *PPEB; typedef struct _GDI_TEB_BATCH { ULONG Offset; ULONG HDC; ULONG Buffer[0x136]; } GDI_TEB_BATCH, *PGDI_TEB_BATCH; typedef struct _NT_TEB { NT_TIB Tib; // 00h PVOID EnvironmentPointer; // 1Ch CLIENT_ID Cid; // 20h PVOID ActiveRpcInfo; // 28h PVOID ThreadLocalStoragePointer; // 2Ch PPEB Peb; // 30h ULONG LastErrorValue; // 34h ULONG CountOfOwnedCriticalSections; // 38h PVOID CsrClientThread; // 3Ch PVOID Win32ThreadInfo; // 40h ULONG Win32ClientInfo[0x1F]; // 44h PVOID WOW32Reserved; // C0h ULONG CurrentLocale; // C4h ULONG FpSoftwareStatusRegister; // C8h PVOID SystemReserved1[0x36]; // CCh PVOID Spare1; // 1A4h LONG ExceptionCode; // 1A8h ULONG SpareBytes1[0x28]; // 1ACh PVOID SystemReserved2[0xA]; // 1D4h GDI_TEB_BATCH GdiTebBatch; // 1FCh ULONG gdiRgn; // 6DCh ULONG gdiPen; // 6E0h ULONG gdiBrush; // 6E4h CLIENT_ID RealClientId; // 6E8h PVOID GdiCachedProcessHandle; // 6F0h ULONG GdiClientPID; // 6F4h ULONG GdiClientTID; // 6F8h PVOID GdiThreadLocaleInfo; // 6FCh PVOID UserReserved[5]; // 700h PVOID glDispatchTable[0x118]; // 714h ULONG glReserved1[0x1A]; // B74h PVOID glReserved2; // BDCh PVOID glSectionInfo; // BE0h PVOID glSection; // BE4h PVOID glTable; // BE8h PVOID glCurrentRC; // BECh PVOID glContext; // BF0h NTSTATUS LastStatusValue; // BF4h UNICODE_STRING StaticUnicodeString; // BF8h WCHAR StaticUnicodeBuffer[0x105]; // C00h PVOID DeallocationStack; // E0Ch PVOID TlsSlots[0x40]; // E10h LIST_ENTRY TlsLinks; // F10h PVOID Vdm; // F18h PVOID ReservedForNtRpc; // F1Ch PVOID DbgSsReserved[0x2]; // F20h ULONG HardErrorDisabled; // F28h PVOID Instrumentation[0x10]; // F2Ch PVOID WinSockData; // F6Ch ULONG GdiBatchCount; // F70h ULONG Spare2; // F74h ULONG Spare3; // F78h ULONG Spare4; // F7Ch PVOID ReservedForOle; // F80h ULONG WaitingOnLoaderLock; // F84h PVOID StackCommit; // F88h PVOID StackCommitMax; // F8Ch PVOID StackReserve; // F90h PVOID MessageQueue; // ??? } NT_TEB, *PNT_TEB; Update (11 Februrary 2003) ====== In Peter Huang whitepaper of compiler security optimization : http://members.rogers.com/yinrong/articles/Prevent.htm For the past few days, I have done a few experiments and some research on ways to prevent the buffer overflow exploitation. I believe the following compiler option (if implemented and used) should make the exploitation of stack buffer overflow by "jmp esp" method impossible (as far as I know) to execute a piece of malicious code on Intel-Inside PC. The buffer vulnerability, if exists, still overflows to be trapped by this mechanism if being pumped too much. For a function like: Void testVoid (void) { do something here and the stack overflows such as sprintf in ssnetlib.dll in SQL 2000 server (pre-SP3) } The compiler inserts the following opcodes before "ret". Mov ecx, [esp] Cmp word ptr [ecx], 0FFE4h ; check for "jmp esp" Jnz realRet Mov byte ptr [esp+4], 0cch ; stack has been overflowed and function ; will breakpoint (no need to go on anyway) RealRet: ret For a function like: Void testParametered(int I, int i2, int i3 ...) { do something here and the stack overflows such as sprintf in ssnetlib.dll in SQL 2000 server (pre-SP3) } The compiler inserts the following opcode before "ret" Mov byte ptr[esp+4], 0cch ; if return address is not overwritten, ; then this will be discarded by opcode such as ; "add esp, xx" down the execution path. ; Otherwise, the exploitation code injected onto the stack, ; which happens before this move, will ; cause a breakpoint trap to happen due to the ; "jmp esp" method exploited by the buffer overflow. Ret Normally, both above functions continue normally. However, if the stack is overflowed due to some software bugs under exploitation attack, then the mechanism listed above will generate breakpoint trap and the execution of malicious code is stopped. Unfortunately, it will definitely increase the executable image size and some runtime hit (should not be much because of the cache lines are still valid etc.). Some might think that this creates some false illusion of security. However, it does dam the overflow to what has been overflowed and prevent the overflow from spreading so that we do not have to hear others say:" dam it".