The DUBNIUM campaign in December involved one exploit in-the-wild that affected Adobe Flash Player. In this blog, we’re going to examine the technical details of the exploit that targeted vulnerability CVE-2015-8651. For more details on this vulnerability, see Adobe Security Bulletin APSB16-01.
Note that Microsoft Edge on Windows 10 was protected from this attack due to the mitigations introduced into the browser.
Adobe Flash Player version checks
The nature of the vulnerability is an integer overflow, and the exploit code has quite extensive subroutines in it. It tries to cover versions of the player from 11.x to the most recent version at the time of the campaign, 18.104.22.168.
The earliest version of Adobe Flash Player 11.x was released in October 2011 (22.214.171.124) and the last version of Adobe Flash Player 10.x was released in June 2013 (10.3.183.90). This doesn’t necessarily mean the exploit existed from 2011 or 2013, but it again demonstrates the broad target the exploit tries to cover.
Figure 1 Version check for oldest Flash Player the exploit targets
Mainly we focused our analysis upon the function named qeiofdsa, as the routine covers any Adobe Flash player version since 126.96.36.199 (released on September 21, 2015).
Figure 2 Version check for latest Flash Player the exploit supports
Why is this version of Flash Player so important? Because that is the release which had the latest Vector length corruption hardening applied at the time of the incident. The original Vector length hardening came with 188.8.131.52 and it is well explained in the Security @ Adobe bloghttps://blogs.adobe.com/security/2015/12/community-collaboration-enhances-flash.html.
The Vector object from Adobe Flash Player can be used as a corruption target to acquire read or write (RW) primitives.
This object has a very simple object structure and predictable allocation patterns without any sanity checks on the objects. This made this object a very popular target for exploitation for recent years. There were a few more bypasses found after that hardening, and 184.108.40.206 had another bypass hardening. The exploit uses a new exploitation method (ByteArray length corruption) since this new version of Adobe Flash Player.
Note, however, that with new mitigation from Adobe released after this incident, the ByteArray length corruption method no longer works.
To better understand the impact of the mitigations on attacker patterns, we compared exploit code line counts for the pdfsajoe routine, which exploits Adobe Flash Player versions earlier than 220.127.116.11, to the qeiofdsa routine, which exploits versions after 18.104.22.168. We learned that pdfsajoe has 139 lines of code versus qeiofdsa with 5,021.
While there is really no absolute way to measure the impact and line code alone is not a standard measurement, we know that in order to target the newer versions of Adobe Flash Player, the attacker would have to write 36 more times the lines of code.
|Vulnerable Flash Player version
||22.214.171.124 and up
||No latest Vector mitigations
||Latest Vector mitigations applied
|Lines of attack code
Table 1 Before and after Vector mitigation
This tells us a lot about the importance of mitigation and the increasing cost of exploit code development. Mitigation in itself doesn’t fix existing vulnerabilities, but it is definitely raising the bar for exploits.
Heap spraying and vulnerability triggering
The exploit heavily relies on heap spraying. Among heap spraying of various objects, the code from Figure 3 shows the code where the ByteArray objects are sprayed. This ByteArray has length of 0x10. These sprayed objects are corruption targets.
Figure 3 Heap-spraying code
The vulnerability lies in the implementation of fast memory opcodes. More detailed information on the usage of fast memory opcodes are available in the Faster byte array operations with ASC2 article at the Adobe Developer Center.
After setting up application domain memory, the code can use avm2.intrinsics.memory. The package provides various methods including li32 and si32 instructions. The li32 can be used to load 32bit integer values from fast memory and si32 can be used to store 32bit integer values to fast memory. These functions are used as methods, but in the AVM2 bytecode level, they are opcode themselves.
Figure 4 Setting up application domain memory
Due to the way these instructions are implemented, the out-of-bounds access vulnerability happens (Figure 5). The key to this vulnerability is the second li32 statement just after first li32 one in each IF statement. For example, from the li32((_local_4+0x7FEDFFD8)) statement, the _local_4+0x7FEDFFD8 value ends up as 4 after integer overflow. From the just-in-time (JIT) level, the range check is only generated for this li32 statement, skipping the range check JIT code for the first li32 statement.
Figure 5 Out-of-bounds access code using li32 instructions
We compared the bytecode level AVM2 instructions with the low-level x86 JIT instructions. Figure 6 shows the comparisons and our findings. Basically two li32 accesses are made and the JIT compiler optimizes length check for both li32 instructions and generates only one length check. The problem is that integer overflow happens and the length check code becomes faulty and allows bypasses of ByteArray length restrictions. This directly ends with out-of-bounds RW access of the process memory. Historically, fast memory implementation suffered range check vulnerabilities (CVE-2013-5330, CVE-2014-0497). The Virus Bulletin 2014 paper by Chun Feng and Elia Florio,Ubiquitous Flash, ubiquitous exploits, ubiquitous mitigation (PDF download), provides more details on other old but similar vulnerabilities.
Figure 6 Length check confusion
Using this out-of-bounds vulnerability, the exploit tries to locate heap-sprayed objects.
These are the last part of memory sweeping code. We counted 95 IF/ELSE statements that sweep through memory range from ba+0x121028 to ba+0x17F028 (where ba is the base address of fast memory), which is 0x5E000 (385,024) byte size. Therefore, these memory ranges are very critical for this exploit’s successful run.
Figure 7 End of memory sweeping code
Figure 8 shows a crash point where the heap spraying fails. The exploit heavily relies on a specific heap layout for successful exploitation, and the need for heap spraying is one element that makes this exploit unreliable.
Figure 8 Out-of-bounds memory access
This exploit uses a corrupt ByteArray.length field and uses it as RW primitives (Figure 9).
Figure 9 Instruction si32 is used to corrupt ByteArray.length field
After ByteArray.length corruption, it needs to determine which ByteArray is corrupt out of the sprayed ByteArrays(Figure 10).
Figure 10 Determining corrupt ByteArray
The following shows various RW primitives that this exploit code provides. Basically these extensive lists of methods provide functions to support different application and operating system flavors.
Figure 11 RW primitives
For example, the read32x86 method can be used to read an arbitrary process’s memory address on x86 platform. The cbIndex variable is the index into the bc array which is an array of the ByteArray type. The bc[cbIndex] is the specific ByteArray that is corrupted through the fast memory vulnerability. After setting virtual address as position member, it uses the readUnsignedInt method to read the memory value.
Figure 12 Read primitive
The same principle applies to the write32x86 method. It uses the writeUnsignedInt method to write to arbitrary memory location.
Figure 13 Write primitive
Above these, the exploit can perform a slightly complex operation like reading multiple bytes using the readBytesmethod.
Figure 14 Byte reading primitive
Function object virtual function table corruption
Just after acquiring the process’s memory RW ability, the exploit tries to get access to code execution. This exploit uses a very specific method of corrupting a Function object and using the apply and call methods of the object to achieve shellcode execution. This method is similar to the exploit method that was disclosed during the Hacking Team leak. Figure 15 shows how the Function object’s virtual function table pointer (vptr) is acquired through a leaked object address, and low-level object offset calculations are performed. The offsets used here are relevant to the Adobe Flash Player’s internal data structure and how they are linked together in the memory.
Figure 15 Resolving Function object vptr address
This leaked virtual function table pointer is later overwritten with a fake virtual function table’s address. The fake virtual function table itself is cloned from the original one and the only pointer to apply method is replaced with the VirtualProtect API. Later, when the apply method is called upon the dummy function object, it will actually call the VirtualProtect API with supplied arguments – not the original empty call body. The supplied arguments are pointing to the memory area that is used for temporary shellcode storage. The area is made read/write/executable (RWX) through this method.
Figure 16 Call VirtualProtect through apply method
Once the RWX memory area is reserved, the exploit uses the call method of the Function object to perform further code execution. It doesn’t use the apply method because it no longer needs to pass any arguments. Calling the call method is also simpler (Figure 17).
Figure 17 Shellcode execution through call method
This shellcode-running routine is highly modularized and you can actually use API names and arguments to be passed to the shellcode-running utility function. This makes shellcode building and running very extensible. Again, this method has close similarity with the code found with the Adobe Flash exploit leaked during the Hacking Team information leak in July 2015.
Figure 18 Part of shellcode call routines
Note that the exploit’s method of using the corrupted Function object virtual table doesn’t work on Microsoft Edge anymore as it has additional mitigation against these kinds of attacks.
With this exploit, shellcode is not just contiguous memory area, but various shellcodes are called through separate call methods. As you can see from this exploit, we are observing more exploits operate without return-oriented programming (ROP) chains. We can track these calls by putting a breakpoint on the native code that performs the ActionScript call method. For example, the disassembly in Figure 19 shows the code that calls theInternetOpenUrlA API call.
Figure 19 InternetOpenUrlA 1st download
This call only retrieves some portion of a portable executable (PE) file’s header, but not the whole file. It will do another run of the InternetOpenUrlA API call to retrieve the remaining body of the payload. This is most likely a trick to confuse analysts who will look for a single download session for payloads.
Figure 20 InternetOpenUrlA 2nd download
With the analysis of the Adobe Flash Player-targeting exploit used by DUBNIUM last December, we learned they are using highly organized exploit code with extensive support of operating system flavors. However, some functionalities for some operating system are not yet implemented. For example, some 64-bit support routines had an empty function inside them.