CERT Basic Fuzzing Framework (BFF) 2.8 ===== Change Log ===== See the NEWS file for changes ===== Requirements ===== The UbuFuzz VM requires VMWare Workstation 7 or later, or compatible virtualization software. If you are using virtualization software other than VMWare, then you must create your own virtual machine that uses the existing vmdk disk file. The OS X installer requires Snow Leopard or later. Xcode Tools is recommended, which provides libgmalloc (Guard Malloc). ===== Demo Quick Start ===== UbuFuzz: 1) Unzip BFF-2.8.zip to c:\fuzz or another folder to be shared as "fuzz" 2) Unzip UbuFuzz-2.8.zip 3) Open UbuFuzz.vmx 4) Create a VM snapshot 5) Power on the VM If you do not wish to use a shared folder, simply remove ~/bff and unzip BFF-2.8.zip to the ~/bff directory. OS X: 1) Install BFF 2) Run BFF 3) Run ./batch.sh To check the progress of the fuzzing campaign, run: ~/bff/quickstats.sh Look in ~/results to investigate the fuzzing results. ===== To use BFF without UbuFuzz ===== Please refer to the INSTALL file. ===== Additional information ===== The configuration for VMWare may prevent virtual machines from utilizing shared folders by default. You may need to manually enable shared folders for the VM after opening the VMX. If you chose to unzip scripts.zip to a folder other than c:\fuzz, then you will need to modify the properties of the shared folder in VMWare to point to the new location of the files. Alternatively, if you may unzip the BFF scripts into ~/bff in the guest OS if you do not wish to use a dhared folder. The fuzzing virtual machine is preconfigured to automatically begin a fuzzing run on several image format decoders provided by ImageMagick's "convert" program. An old (5.2.0) version of ImageMagick is preloaded onto the VM. ImageMagick was built with debug symbols as well as non-optimized. This makes gdb provide more useful crash reports. ImageMagick was configured using the following command: CFLAGS="-g -O0" ./configure --without-x It is recommended to perform fuzzing with virtual network disabled to prevent the fuzzing target and/or debugger from communicating unnecessarily. ===== Analyzing results ===== When the fuzzing run encounters a crash, BFF will analyze the details of the crash. This involves capturing stderr, gdb, valgrind, and callgrind output. The gdb output contains several pieces of information, including the memory map, signal information, backtrace, registers, disassembly, as well as output from the CERT Triage Tools, which indicates possible exploitability of the crash. On the OS X platform, CrashWrangler is used instead of gdb. By looking at the backtrace, BFF will keep track of which test cases cause unique crashes. Each unique crash will be placed in the configured output directory. results/ |-- <target_id> | |-- crashers | | |-- <exploitability> | | | |-- <crash_id> | | | | |-- <analysis_tools>.stderr | | | | |-- minimizer_log.txt | | | | |-- sf_<seedfile_md5>-<bff_iteration_number>-minimized.<seedfile_ext> | | | | |-- sf_<seedfile_md5>-<bff_iteration_number>-minimized.<seedfile_ext>.callgrind | | | | |-- sf_<seedfile_md5>-<bff_iteration_number>-minimized.<seedfile_ext>.callgrind.annotated | | | | |-- sf_<seedfile_md5>-<bff_iteration_number>-minimized.<seedfile_ext>.callgrind.calltree | | | | |-- sf_<seedfile_md5>-<bff_iteration_number>-minimized.<seedfile_ext>.drillresults | | | | |-- sf_<seedfile_md5>-<bff_iteration_number>-minimized.<seedfile_ext>.gdb | | | | |-- sf_<seedfile_md5>-<bff_iteration_number>-minimized.<seedfile_ext>.stderr | | | | |-- sf_<seedfile_md5>-<bff_iteration_number>-minimized.<seedfile_ext>.valgrind | | | | |-- sf_<seedfile_md5>-<bff_iteration_number>.<seedfile_ext> | | | | `-- sf_<seedfile_md5>.<seedfile_ext> | |-- seedfiles | |-- sf_<seedfile1_md5>.<seedfile1_ext> | |-- sf_<seedfile2_md5>.<seedfile2_ext> `-- uniquelog.txt The UbuFuzz startup script, batch.sh, looks for the bff.py script in the same directory, which by default is a soft link to the VMWare shared folder /mnt/hgfs/fuzz. BFF will copy its configuration to results/<target_id>/bff.yaml, and will log messages of level INFO or higher into log/bff.log. The config file copied here is just for recording purposes. Additionally the following subdirectories are created in the results dir, in the <target_id> subdirectory: * crashers: Contains a subdir for each uniquely-crashing test case and its analyzed results * seeds: Contains the original seedfiles as well as logs specific to that seedfile Other files of note: results/<target_id>/uniquelog.txt a log file that tracks the unique crashers found during the run The "results/<target_id>/crashers" directory will contain the uniquely-crashing test cases. The variants that have crashed the target application will be stored here with the BFF iteration number appended to the seed file name. For each uniquely-crashing case, there will also be a .stderr, .gdb, .drillresults, .callgrind and .valgrind file that contains the stderr, gdb, drillresults, callgrind and valgrind output for that case, respectively. The CERT BFF minimizes crashers. In other words, each crashing test case will have a number of bytes that have been modified from the seed file. When a crasher is minimized, a test case is generated with a minimal number of bytes that have changed from the seed file. The minimized test case will have "-minimized" inserted into the filename before the extension. Also, a minimizer_log.txt file is created containing a log of the minimization process. This is used by tools/minimizer_plot.py to produce a chart showing the minimizer progress. Aside from minimizing to the seed file, BFF also includes the ability to minimize a crashing testcase to a string pattern. This will show which bytes of a file may be altered without affecting the crash. Due to performance issues, the minimize-to-string option is disabled by default. However, the minimizer can be run in a standalone mode. Say you have a crashing test case, but you really need to get it to a proof-of-concept exploit. The problem is when you load the crash into your debugger you can't easily tell which registers, stack values, or memory locations are under your control. But what if you could change the crashing test case so that it had only the bytes required to cause that crash, and the rest were all masked out with a fixed value, say "x" (0x78)? Then you'd know that if you saw EIP=0x78787878, you may already be a winner. The minimize-to-string option does just that. To get command-line usage of the minimizer, run: tools/minimize.py --help To minimize a crashing testcase to the Metasploit string pattern, run: tools/minimize.py --stringmode <crashing_testcase> When minimizing to the Metasploit pattern, FOE will use the resulting byte map to create an additional minimized file that uses a string of 'x' characters. Note that this file is not guaranteed to produce the same crash as the original string minimization. Metasploit pattern enumeration: Especially with larger files, you may notice that the Metasploit pattern repeats several times over the length of a Metasploit-minimized crasher. Given any particular dword, it may not be obvious which instance is the one that you are dealing with. This is where the tools/mtsp_enum.py script comes in handy. For example, let's say that you have a crasher.doc were EIP = "Aa0A" If you run: tools/mtsp_enum.py Aa0A crasher.doc You will end up with a file called crasher-enum.doc. With this file, every instance of the byte pattern "Aa0A" will be replaced with a unique, incrementing replacement. For example, "0a0A", "1a0A", "2a0A", etc. Now when you open crasher-enum.doc, you could for example get EIP = "5a0A". If you search for that pattern in the file, there should be only once instance of it. Note that you can use a search pattern of any length and you can also search for hex values. For example: "\x01\x02\x03\x04" ===== Analysis tools ===== The analysis directory contains a few tools for analyzing the results of a fuzz run. Try 'python <script> --help' for detailed usage options. tools/bff_stats.py generates a concise summary of the fuzz run results so far, including how many times each unique crash was seen, the first seed number it was seen at, the most recent seed number it was seen, and the bitwise and bytewise Hamming Distance from the original seedfile for the minimized testcase. tools/callsim.py can display crashes clustered by the similarity of their called functions. This analysis is based on the idea that crashes with similar call history are likely to be related even if they result in unique crash hashes. The resulting clusters of crashes can be useful in deciding which crashes to investigate first given a large number of crashes. The source data for this analysis is callgrind output generated for each crash. tools/create_crasher_script.py will generate a shell script that in turn can be used to regenerate all the test cases for a given crash id. The use of the '--destination' option is highly recommended. tools/minimizer_plot.py plots the minimizer data for a given crash, showing how the minimizer tunes its parameters as it progresses in order to find the optimal minimized test case. tools/drillresults.py will search for crashing test cases that are more easily exploitable than the others. It searches based on the type of crash as well as whether the faulting address matches patterns in the fuzzed file. tools/repro.py will launch the specified application using the same command-line parameters as configure for the fuzzing campaign. This can be used to test crashing testcases interactively. tools/zipdiff.py can be used to compare zip-based files. tools/debuggerfile.py will parse GDB output to create a hash that uniquely identifies the crash. tools/copycrashers.py to collect all of the crashing cases from a campaign. By default it will copy all of the uniquely-crashing test cases to the "seedfiles" directory, which should be empty. ===== Updating BFF ===== BFF includes a script to update itself using the code present on GitHub <https://github.com/CERTCC-Vulnerability-Analysis/certfuzz> tools/updatebff.py This script will first attempt to use GitPython to retrieve updates. If this package is not available, it will fall back to using the system-wide installed git binary. Finally, if git is not available, it will fall back to using python urllib. Note that urllib does not support retrieving HTTPS URIs through a proxy. If your machines are behind a proxy, you must have either GitPython or git installed to use the updatebff.py script. ===== Fuzzing on your own ===== When the UbuFuzz VM is powered on, it will automatically execute the batch.sh script in the VMWare shared folder. In order to power on the virtual machine without it beginning a fuzzing campaign, you should rename batch.sh. This will allow you to power on the virtual machine to install the target software. The first step to beginning a fuzzing run is to obtain and install the target software. Usually this process will involve downloading the application source code and compiling a debug version of it. Rather than using Ubuntu's package repository, this will ensure that you will be fuzzing the latest version of the target application. It will also give you the ability to have debug symbols, as well as the ability to use a non-optimized build if you like. When compiling the target application, be sure that the "-g" flag is used with your compiler. This ensures that debug information is added to the target application. Create a new snapshot of the VM after the target software has been installed. The configs/bff.yaml file contains all of the parameters for the fuzzing run. This file must be edited to suit the software that you will be fuzzing. The bff.yaml file is annotated and should be relatively self-explanatory. The default bff.yaml file will start a fuzzing run that invokes: ~/convert $SEEDFILE /dev/null This will use ImageMagick's "convert" program to process the seed file, outputting to /dev/null. Because BFF will mangle the seed file, this fuzzing run will exercise ImageMagick's decoding capabilities. You should choose a command line that efficiently exercies the capabilities of the application that you wish to test. Ideally, this will involve a command-line application that terminates immediately upon completion. In such cases, the fuzzing run will be CPU-bound, as BFF will invoke the application as fast as it can. Applications that have a GUI can still be fuzzed using this framework. However, in most cases, each invocation of the target application will need to wait for the specified timeout to elapse before the process is terminated. BFF analyzes gdb (or CrashWrangler on OS X) backtraces to determine uniquely-crashing testcases. The default setting is to hash the last five code locations in the backtrace to determine a hash for the crash. We have found this value to be effective for most applications, however you can adjust the value to fit your needs. There are several options within bff.yaml related to application timeouts. "runtimeout" is the maximum time that BFF will allow a single invocation of the target application to run before terminating it. In the case of the "convert" component of ImageMagick, it is reasonable to expect the program to finish within a few seconds. Depending on the application you are fuzzing, this value will need to be adjusted. Especially with GUI applications, the goal is to allow the application to run long enough to process the input file, but not so long that the application is sitting idle for an amount of time for each invocation. Once you have configured the options in bff.yaml, power on the UbuFuzz VM. It should begin fuzzing automatically. If the target is a GUI application, you should see the application launch repeatedly as it is being fuzzed. If it is a command-line application, you should notice increased CPU usage in the htop window. The first iteration that BFF executes will also display stdout and stderr to the terminal window to let you see if something is obviously wrong. If you need to tweak your settings, such as by using a smaller progtimeout value to improve throughput, you can edit the bff.yaml file and restore the VM to its previous snapshot. Alternatively, you can run ~/bff/reset_bff.sh and restart ~/bff/batch.sh to start fuzzing again.