We recently gave a talk at the ITWeb Security Summit entitled "Offense Oriented Defence". The talk was targeted at defenders and auditors, rather then hackers (the con is oriented that way), although it's odd that I feel the need to apologise for that ;)
The talks primary point, was that by understanding how attackers attack, more innovative defences can be imagined. The corollary was that common defences, in the form of "best practise" introduce commonality that is more easily exploited, or at least degrade over time as attackers adapt. Finally, many of these "security basics" are honestly hard, and we can't place the reliance on them we'd hoped. But our approach doesn't seem to want to acknowledge the problem, and much like an AA meeting, it's time we recognise the problem.
If you had to look at the average security strategy or budget items, you often end up with a list containing a couple of these:
But, the truth is many of these items don't actually block attacks, or the few that do, don't really counter the common bypassed used to side-step them. For example:
The current place we seem to be in is akin to having everyone build a wall. Attackers get to evaluate the wall, figure out how to get over it, and add to their capability (i.e. get a longer rope). But once they have a longer rope, they can use it over and over again, and against more than one wall. So attackers, who are quite good at sharing, get to keep building their tool chain, while all defenders can do it to keep building a higher wall, and maintaining the increasingly untenable structure. By understanding how attackers attack, we can break out of this and try more innovative approaches.
The talk is illustrated with four broad examples: Passwords, Patches, Anti-Virus and DMZs. For each, the belief around specific configurations is discussed, and how those don't stand up to how attackers actually attack. For example, the way AV's believed to work doesn't seem to correspond with how easy they are to bypass, or the common configuration of standard password controls such as lockout, don't seem to take into account horizontal brute-force attacks.
The point I want to make here is somewhat subtle; if you walk away thinking I've described new attacks, then you've missed it, if you think I'm recommending "the basics" then you've missed it. Truthfully, maybe it's just that I didn't make it very well ... decide for yourself, here are the slides:
New types of mobile applications based on Trusted Execution Environments (TEE) and most notably ARM TrustZone micro-kernels are emerging which require new types of security assessment tools and techniques. In this blog post we review an example TrustZone application on a Galaxy S3 phone and demonstrate how to capture communication between the Android application and TrustZone OS using an instrumented version of the Mobicore Android library. We also present a security issue in the Mobicore kernel driver that could allow unauthorised communication between low privileged Android processes and Mobicore enabled kernel drivers such as an IPSEC driver.
Mobicore OS :
The Samsung Galaxy S III was the first mobile phone that utilized ARM TrustZone feature to host and run a secure micro-kernel on the application processor. This kernel named Mobicore is isolated from the handset's Android operating system in the CPU design level. Mobicore is a micro-kernel developed by Giesecke & Devrient GmbH (G&D) which uses TrustZone security extension of ARM processors to create a secure program execution and data storage environment which sits next to the rich operating system (Android, Windows , iOS) of the Mobile phone or tablet. The following figure published by G&D demonstrates Mobicore's architecture :
The security critical applications that run inside Mobicore OS are referred to as trustlets and are developed by third-parties such as banks and content providers. The trustlet software development kit includes library files to develop, test and deploy trustlets as well as Android applications that communicate with relevant trustlets via Mobicore API for Android. Trustlets need to be encrypted, digitally signed and then remotely provisioned by G&D on the target mobile phone(s). Mobicore API for Android consists of the following 3 components:
1) Mobicore client library located at /system/lib/libMcClient.so: This is the library file used by Android OS or Dalvik applications to establish communication sessions with trustlets on the secure world
2) Mobicore Daemon located at /system/bin/mcDriverDaemon: This service proxies Mobicore commands and responses between NWd and SWd via Mobicore device driver
3) Mobicore device driver: Registers /dev/mobicore device and performs ARM Secure Monitor Calls (SMC) to switch the context from NWd to SWd
The source code for the above components can be downloaded from Google Code. I enabled the verbose debug messages in the kernel driver and recompiled a Samsung S3 kernel image for the purpose of this analysis. Please note that you need to download the relevant kernel source tree and stock ROM for your S3 phone kernel build number which can be found in "Settings->About device". After compiling the new zImage file, you would need to insert it into a custom ROM and flash your phone. To build the custom ROM I used "Android ROM Kitchen 0.217" which has the option to unpack zImage from the stock ROM, replace it with the newly compiled zImage and pack it again.
1) Android application calls mcOpenDevice() API which cause the Mobicore Daemon (/system/bin/mcDriverDaemon) to open a handle to /dev/mobicore misc device.
2) It then allocates a "Worlds share memory" (WSM) buffer by calling mcMallocWsm() that cause the Mobicore kernel driver to allocate wsm buffer with the requested size and map it to the user space application process. This shared memory buffer would later be used by the android application and trustlet to exchange commands and responses.
3) The mcOpenSession() is called with the UUID of the target trustlet (10 bytes value, for instance : ffffffff000000000003 for PlayReady DRM truslet) and allocate wsm address to establish a session with the target trustlet through the allocated shared memory.
4) Android applications have the option to attach additional memory buffers (up to 6 with maximum size of 1MB each) to the established session by calling mcMap() API. In case of PlayReady DRM trustlet which is used by the Samsung VideoHub application, two additional buffers are attached: one for sending and receiving the parameters and the other for receiving trustlet's text output.
5) The application copies the command and parameter types to the WSM along with the parameter values in second allocated buffer and then calls mcNotify() API to notify the Mobicore that a pending command is waiting in the WSM to be dispatched to the target trustlet.
6) The mcWaitNotification() API is called with the timeout value which blocks until a response received from the trustlet. If the response was not an error, the application can read trustlets' returned data, output text and parameter values from WSM and the two additional mapped buffers.
7) At the end of the session the application calls mcUnMap, mcFreeWsm and mcCloseSession .
The Mobicore kernel driver is the only component in the android operating system that interacts directly with Mobicore OS by use of ARM CPU's SMC instruction and Secure Interrupts . The interrupt number registered by Mobicore kernel driver in Samsung S3 phone is 47 that could be different for other phone or tablet boards. The Mobicore OS uses the same interrupt to notify the kernel driver in android OS when it writes back data.
Analysis of a Mobicore session:
There are currently 5 trustlets pre-loaded on the European S3 phones as listed below:
shell@android:/ # ls /data/app/mcRegistry
The 07010000000000000000000000000000.tlbin is the "Content Management" trustlet which is used by G&D to install/update other trustlets on the target phones. The 00060308060501020000000000000000.tlbin and ffffffff000000000000000000000003.tlbin are DRM related truslets developed by Discretix. I chose to analyze PlayReady DRM trustlet (ffffffff000000000000000000000003.tlbin), as it was used by the Samsung videohub application which is pre-loaded on the European S3 phones.
The videohub application dose not directly communicate with PlayReady trustlet. Instead, the Android DRM manager loads several DRM plugins including libdxdrmframeworkplugin.so which is dependent on libDxDrmServer.so library that makes Mobicore API calls. Both of these libraries are closed source and I had to perform dynamic analysis to monitor communication between libDxDrmServer.so and PlayReady trustlet. For this purpose, I could install API hooks in android DRM manager process (drmserver) and record the parameter values passed to Mobicore user library (/system/lib/libMcClient.so) by setting LD_PRELOAD environment variable in the init.rc script and flash my phone with the new ROM. I found this approach unnecessary, as the source code for Mobicore user library was available and I could add simple instrumentation code to it which saves API calls and related world shared memory buffers to a log file. In order to compile such modified Mobicore library, you would need to the place it under the Android source code tree on a 64 bit machine (Android 4.1.1 requires 64 bit machine to compile) with 30 GB disk space. To save you from this trouble, you can download a copy of my Mobicore user library from here. You need to create the empty log file at /data/local/tmp/log and replace this instrumented library with the original file (DO NOT FORGET TO BACKUP THE ORIGINAL FILE). If you reboot the phone, the Mobicore session between Android's DRM server and PlayReady trustlet will be logged into /data/local/tmp/log. A sample of such session log is shown below:
The content and address of the shared world memory and two additional mapped buffers are recorded in the above file. The command/response format in wsm buffer is very similar to APDU communication in smart card applications and this is not a surprise, as G&D has a long history in smart card technology. The next step is to interpret the command/response data, so that we can manipulate them later and observe the trustlet behavior. The trustlet's output in text format together with inspecting the assembly code of libDxDrmServer.so helped me to figure out the PlayReady trustlet command and response format as follows:
client command (wsm) : 08022000b420030000000001000000002500000028023000300000000500000000000000000000000000b0720000000000000000
client parameters (mapped buffer 1): 8f248d7e3f97ee551b9d3b0504ae535e45e99593efecd6175e15f7bdfd3f5012e603d6459066cc5c602cf3c9bf0f705b
trustlet response (wsm):08022000b420030000000081000000002500000028023000300000000500000000000000000000000000b0720000000000000000
trustltlet text output (mapped buffer 2):
SRVXInvokeCommand command 1000000 hSession=320b4
SRVXInvokeCommand. command = 0x1000000 nParamTypes=0x25
SERVICE_DRM_BBX_SetKeyToOemContext - pPrdyServiceGlobalContext is 32074
SERVICE_DRM_BBX_SetKeyToOemContext iExpectedSize match real size=48
SERVICE_DRM_BBX_SetKeyToOemContext preparing local buffer DxDecryptAsset start - iDatatLen=32, pszInData=0x4ddf4 pszIntegrity=0x4dde4
DxDecryptAsset calling Oem_Aes_SetKey DxDecryptAsset
calling DRM_Aes_CtrProcessData DxDecryptAsset
calling DRM_HMAC_CreateMAC iDatatLen=32 DxDecryptAsset
after calling DRM_HMAC_CreateMAC DxDecryptAsset
By mapping the information disclosed in the trustlet text output to the client command the following format was derived:
08022000 : virtual memory address of the text output buffer in the secure world (little endian format of 0x200208)
b4200300 : PlayReady session ID
00000001: Command ID (0x1000000)
00000000: Error code (0x0 = no error, is set by truslet after mcWaitNotification)
25000000: Parameter type (0x25)
28023000: virtual memory address of the parameters buffer in the secure world (little endian format of 0x300228)
30000000: Parameters length in bytes (0x30, encrypted key length)
05000000: encryption key type (0x5)
The trustlet receives client supplied memory addresses as input data which could be manipulated by an attacker. We'll test this attack later. The captured PlayReady session involved 18 command/response pairs that correspond to the following high level diagram of PlayReady DRM algorithm published by G&D. I couldn't find more detailed specification of the PlayReady DRM on the MSDN or other web sites. But at this stage, I was not interested in the implementation details of the PlayReady schema, as I didn't want to attack the DRM itself, but wanted to find any exploitable issue such as a buffer overflow or memory disclosure in the trustlet.
An attacker would need to know the "sequence number" of an already established netlink connection between a kernel component such as IPSEC and Mobicore driver in order to exploit this vulnerability. This sequence numbers were incremental starting from zero but currently there is no kernel component on the Samsung phone that uses the Mobicore API, thus this issue was not a high risk. We notified the vendor about this issue 6 months ago but haven't received any response regarding the planned fix. The following figures demonstrate exploitation of this issue from an Android unprivileged process :
|0||Memory address of the mapped output buffer in trustlet process (original value=0x08022000)||for values<0x8022000 the fuzzer crashed|
values >0x8022000 no errors
|41||memory address of the parameter mapped buffer in trusltet process (original value=0x28023000)||0x00001000<value<0x28023000 the fuzzer crashed|
value>=00001000 trustlet exits with "parameter refers to secure memory area"
value>0x28023000 no errors
|49||Parameter length (encryption key or certificate file length)||For large numbers the trustlet exits with "malloc() failed" message|
We demonstrated that intercepting and manipulating the worlds share memory (WSM) data can be used to gain better knowledge about the internal workings of Mobicore trustlets. We believe that this method can be combined with the side channel measurements to perform blackbox security assessment of the mobile TEE applications. The context switching and memory sharing between normal and secure world could be subjected to side channel attacks in specific cases and we are focusing our future research on this area.
We're excited to be presenting our Hacking By Numbers Combat course again at Black Hat USA this year. SensePost's resident German haxor dude Georg-Christian Pranschke will be presenting this year's course. Combat fits in right at the top of our course offerings. No messing about, this really is the course where your sole aim is to pwn as much of the infrastructure and applications as possible. It is for the security professional looking to hone their skill-set, or to think like those in Unit 61398. There are a few assumptions though:
These targets come from real life assessments we've faced at SensePost, it's about as real as you can get without having to do the report at the end of it. How it works is that candidates are presented with a specific goal. If the presenter is feeling generous at the time, they may even get a description of the technology. After that, they'll have time to solve the puzzle. Afterwards, there will be a discussion about the failings, takeaways and alternate approaches adopted by the class. The latter is normally fascinating as (as anybody in the industry knows), there are virtually a limitless number of different ways to solve specific problems. This means that even the instructor gets to learn a couple of new tricks (we also have prizes for those who teach them enough new tricks).
In 2012, Combat underwent a massive rework and we presented a virtually new course which went down excellently. We're aiming to do the same this year, and to make it the best Combat course ever. So if you're interested in spending two days' worth of intense thinking solving some fairly unique puzzles and shelling boxen, join us for HBN Combat at BlackHat USA.
Monday morning, raring for a week of pwnage and you see you've just been handed a new assessment, awesome. The problem? It's a mobile assessment and you've never done one before. What do you do, approach your team leader and ask for another assessment? He's going to tell you to learn how to do a mobile assessment and do it quickly, there are plenty more to come.
Now you set out on your journey into mobile assessments and you get lucky, the application that needs to be assessed is an Android app. A few Google searches later and you are feeling pretty confident about this, Android assessments are meant to be easy, there are even a few tools out there that "do it all". You download the latest and greatest version, run it and the app gets a clean bill of health. After all, the tool says so, there is no attack surface; no exposed intents and the permissions all check out. You compile your report, hand it off to the client and a week later the client gets owned through the application... Apparently the backend servers were accepting application input without performing any authentication checks. Furthermore, all user input was trusted and no server side validation was being performed. What went wrong? How did you miss these basic mistakes? After-all, you followed all the steps, you ran the best tools and you ticked all the boxes. Unfortunately this approach is wrong, mobile assessments are not always simply about running a tool, a lot of the time they require the same steps used to test web applications, just applied in a different manner. This is where SensePost's Hacking by numbers: Mobile comes to the fore, the course aims to introduce you to mobile training from the ground up.
The course offers hands-on training, introducing techniques for assessing applications on Android, IOS, RIM and Windows 8. Some of the areas covered include:
On your next mobile assessment you'll be able to do both static and dynamic analysis of mobile applications. You will know where to find those credit card numbers stored on the phone and how to intercept traffic between the application and the backend servers.
The course: Hacking by numbers: Mobile
A few days ago, during one of those nights with the baby crying at 2:00 am and the only thing you can do is to read emails, I realised that Gmail shows the content of compressed files when reading them in Google Docs. As often is the case at SensePost, the "think evil (tm)" came to me and I started to ponder the possibilities of injecting HTML inside the file listing. The idea is actually rather simple. Looking at the file format of a .zip file we see the following:
Every file in the compressed file must have two entries; ZipFileRecord and ZipDirEntry. Both of these entries contain the filename, but only the first one contains the length of filename (it must match the actual length). Our first test case is obvious; if we could modify this name once the file was compressed, would Google sanitise it? Thankfully, the answer is, yes! (go Google!)
As you can see, Google shows the file name inside the compressed file but the tag is displayed with HTML entities. If we then try to see the contents of the file, Google responds by telling us it's not possible to read the content of the file (it's empty) and shows you the file "without formatting" after a few seconds:
Finally, the filename is shown but not sanitised:
Why this is possible?
Remember that the zip format has the name of the compressed files twice. Google uses the first one (ZipFileRecord) for displaying the file names, but in the vulnerable page it uses the second one (ZipDirEntry).
Possible attack vectors
Going back to the 'thinking evil (tm)' mindset, it is now possible to leave a "comprehensive" name in the first entry and inject the malicious payload in the second one. When I first discovered the possibility of doing this, I contacted Google, however, the XSS is in the googleusercontent.com domain, which Google's security team described as a "sandbox" domain (i.e. we aren't injecting into the DOM of google.com) and therefore not worthy of a bounty. Which I accept, if I had to prove usefulness this could be used as part of a simple social engineering attack, for example:
Leading the victim to my phishing site:
Which then proceeds to steals their Google session, or allows the attacker to use BeEF:
Granted, there are simpler ways of achieving the same result. I just wanted to demonstrate how you can use file meta-information for such an attack.