- Into the House of Logic
- Should Reverse Engineering Be Illegal?
- Reverse Engineering Tools and Concepts
- Approaches to Reverse Engineering
- Methods of the Reverser
- Writing Interactive Disassembler (IDA) Plugins
- Decompiling and Disassembling Software
- Decompilation in Practice: Reversing helpctr.exe
- Automatic, Bulk Auditing for Vulnerabilities
- Writing Your Own Cracking Tools
- Building a Basic Code Coverage Tool
- Conclusion
This chapter is from the book
This chapter is from the book
This chapter is from the book
Methods of the Reverser
There are several methods that can be used while reverse engineering software. Each has benefits and each has resource and time requirements. A typical approach uses a mixture of methods when decompiling and examining software. The best method mix depends entirely on your goals. For example, you may first want to run a quick scan of the code for obvious vulnerabilities. Next, you may want to perform a detailed input trace on the user-supplied data. You may not have time to trace each and every path, so you may use complex breakpoints and other tools to speed up the process. What follows is a brief description of several basic methods.
Tracing Input
Input tracing is the most thorough of all methods. First you identify the input points in the code. Input points are places where user-supplied data are being delivered to the program. For example, a call to WSARecvFrom() will retrieve a network packet. This call, in essence, accepts user-supplied data from the network and places it in a buffer. You can set a breakpoint on the input point and single-step trace into the program. Of course, your debugging tools should always include a pencil and paper. You must note each twist and turn in the code path. This approach is very tedious, but it is also very comprehensive.
Although determining all input points takes a great deal of time if you do it by hand, you have the opportunity to note every single code location that makes decisions based on user-supplied data. Using this method you can find very complex problems.
One language that protects against this kind of "look through the inputs" attack is Perl. Perl has a special security mode called taint mode. Taint mode uses a combination of static and dynamic checks to monitor all information that comes from outside a program (such as user input, program arguments, and environment variables) and issues warnings when the program attempts to do something potentially dangerous with that untrusted information. Consider the following script:
#!/usr/bin/perl -T
$username = <STDIN>;
chop $username;
system ("cat /usr/stats/$username");
On executing this script, Perl enters taint mode because of the –T option passed in the invocation line at the top. Perl then tries to compile the program. Taint mode will notice that the programmer has not explicitly initialized the PATH variable, yet tries to invoke a program using the shell anyway, which can easily be exploited. It issues an error such as the following before aborting compilation:
Insecure $ENV{PATH} while running with -T switch at
./catform.pl line 4, <STDIN> chunk 1.
We can modify the script to set the program's path explicitly to some safe value at startup:
#!/usr/bin/perl -T
use strict;
$ENV{PATH} = join ':' => split (" ",<< '__EOPATH__');
/usr/bin
/bin
__EOPATH__
my $username = <STDIN>;
chop $username;
system ("cat /usr/stats/$username");
Taint mode now determines that the $username variable is externally controlled and is not to be trusted. It determines that, because $username may be poisoned, the call to system may be poisoned. It thus gives an other error:
Insecure dependency in system while running with -T switch at ./catform.pl line 9, <STDIN> chunk 1.
Even if we were to copy $username into another variable, taint mode would still catch the problem.
In the previous example, taint mode complains because the variable can use shell magic to cause a command to run. But taint mode does not address every possible input vulnerability, so a clever attacker using our input-driven method can still win.
Advanced dataflow analysis is also useful to help protect against our attack method (or to help carry it out). Static analysis tools can help an analyst (or an attacker) identify all possible input points and to determine which variables are affected from the outside. The security research literature is filled with references discussing "secure information flow" that take advantage of data flow analysis to determine program safety.
Exploiting Version Differences
When you study a system to find weaknesses, remember that the software vendor fixes many bugs in each version release. In some cases the vendor may supply a "hot fix" or a patch that updates the system binaries. It is extremely important to watch the differences between software versions.
The differences between versions are, in essence, attack maps. If a new version of the software or protocol specification is available, then weaknesses or bugs will most certainly have been fixed (if they have been discovered). Even if the "bug fix" list is not published, you can compare the binary files of the older version against the new. Differences can be uncovered where features have been added or bugs have been fixed. These differences thereby reveal important hints regarding where to look for vulnerabilities.
Making Use of Code Coverage
Cracking a computer system is a scientific process just as much as it is an art. In fact, wielding the scientific method gives the attacker an upper hand in an otherwise arbitrary game. The scientific method starts with measurement. Without the ability to measure your environment, how can you possibly draw conclusions about it? Most of the approaches we consider in this text are designed to find programming flaws. Usually (not always), the bugs we find this way are confined to small regions of code. In other words, it's usually the small coding mistakes that we are after. This is one reason that new development tools are very likely to hamper many of the traditional methods of attack. It's easy for a development tool to identify a simple programming error (statically) and compile it out. In a few years, buffer overflows will be obsolete as an attack method.
All the techniques we describe are a form of measurement. We observe the behavior of the program while it is exercised in some way (for example, placed under stress). Strange behavior usually indicates unstable code. Unstable code has a high probability of security weaknesses. Measurement is the key.
Code coverage is an important type of measurement—perhaps the most important. Code coverage is a way of watching a program execute and determining which code paths have been exercised. Many tools are available for code coverage analysis. Code coverage tools do not always require source code. Some tools can attach to a process and gather measurements in real time. For one example, check out the University of Maryland's tool dyninstAPI (created by Jeff Hollingsworth). [7]
As an attacker, code coverage tells you how much work is left to do when you're surveying the landscape. By using coverage analysis you can immediately learn what you have missed. Computer programs are complex, and cracking them is tedious business. It's human nature to skip parts of the code and take shortcuts. Code coverage can show you whether you have missed something. If you skipped that subroutine because it looked harmless, well think again! Code coverage can help you go back and check your work, walking down those dark alleys you missed the first time.
If you are trying to crack software, you most likely start with the user input point. As an example, consider a call to WSARecv(). [8] Using outside-in tracing, you can measure the code paths that are visited. Many decisions are made by the code after user input is accepted. These decisions are implemented as branching statements, such as the conditional branch statements JNZ and JE, in x86 machine code. A code coverage tool can detect when a branch is about to occur and can build a map of each continuous block of machine code. What this means is that you, as the attacker, can instantly determine which code paths you have not exercised during your analysis.
Reverse engineers know that their work is long and tedious. Using code coverage gives the clever reverse engineer a map for tracking progress. Such tracking can keep you sane and can also keep you going when you otherwise might give up without exploring all opportunities.
Code coverage is such an important tool for your bag of tricks that later in the chapter we illustrate how you can build a code coverage tool from scratch. In our example we focus on the x86 assembly language and the Windows XP OS. Our experience leads us to believe that it will be hard for you to find the perfect code coverage tool for your exact needs. Many of the available tools, commercial or otherwise, lack attack-style features and data visualization methods that are important to the attacker.
Accessing the Kernel
Poor access controls on handles opened by drivers can expose a system to attack. If you find a device driver with an unprotected handle, you might be able to run IOCTL commands to the kernel driver. Depending on what the driver supports, you might be able to crash the machine or gain access to the kernel. Any input to the driver that includes memory addresses should be immediately tested by inserting NULL values. Another option is to insert addresses that map to kernel memory. If the driver doesn't perform sanity checking on the user-mode-supplied values, kernel memory may get malformed. If the attack is very clever, global state in the kernel may be modified, altering access permissions.
Leaking Data in Shared Buffers
Sharing buffers is somewhat like sharing food. A restaurant (hopefully) maintains strict rules about where raw meat can be placed. A little raw juice in someone's cooked meal could lead to illness and a lawsuit. A typical program has many buffers. Programs tend to reuse the same buffers over and over, but the questions from our perspective are the following: Will they be cleaned? Are dirty data kept from clean data? Buffers are a great place to start looking for potential data leakage. Any buffer that is used for both public and private data has a potential to leak information.
Attacks that cause state corruption and/or race conditions may be used to cause private data to leak into public data. Any use of a buffer without cleaning the data between uses leads to potential leaks.
Example: The Ethernet Scrubbing Problem
One of us (Hoglund) codiscovered a vulnerability a few years ago that affects potentially millions of ethernet cards worldwide. [9] Ethernet cards use standard chip sets to connect to the network. These chips are truly the "tires" of the Internet. The problem is that many of these chips are leaking data across packets.
The problem exists because data are stored in a buffer on the ethernet microchip. The minimum amount of data that must be sent in an ethernet packet is 66 bytes. This is the minimum frame size. But, many packets that need to be transmitted are actually much smaller than 66 bytes. Examples include small ping packets and ARP requests. Thus, these small packets are padded with data to meet the minimum number of 66 bytes.
The problem? Many chips do not clean their buffers between packets. Thus, a small packet will be padded with whatever was left in the buffer from the last packet. This means that other people's packets are leaking into a potential attack packet. This attack is simple to exploit and the attack works over switched environments. An attack can craft a volley of small packets that solicit a small packet as a reply. As the small reply packets arrive, the attacker looks at the padding data to see other people's packet data.
Of course, some data are lost in this attack, because the first part of every packet is overwritten with the legitimate data for the reply. So, the attacker will naturally want to craft as small a packet as possible to siphon the data stream. Ping packets work well for these purposes, and allow an attacker to sniff cleartext passwords and even parts of encryption keys. ARP packets are even smaller, but will not work as a remote attack. Using ARP packets, an attacker can get TCP ACK numbers from other sessions in the response. This aids in a standard TCP/IP hijacking attack. [10]
Auditing for Access Requirement Screwups
Lack of planning or laziness on the part of software engineers often leads to programs that require administrator or root access to operate. [11] Many programs that were upgraded from older Windows environments to work on Win2K and Windows XP usually require full access to the system. This would be OK except that programs that operate this way tend to leave a lot of world-accessible files sitting around.
Look for directories where user data files are being stored. Ask yourself, are these directories storing sensitive data as well? If so, is the directory permission weak? This applies to the NT registry and to database operations as well. If an attacker replaces a DLL or changes the settings for a program, the attacker might be able to elevate access and take over a system. Under Windows NT, look for open calls that request or create resources with no access restrictions. Excessive access requirements lead to insecure file and object permissions.
Using Your API Resources
Many system calls are known to lead to potential vulnerabilities [Viega and McGraw, 2001]. One good method of attack when reversing is to look for known calls that are problematic (including, for example, the much maligned strcpy()). Fortunately, there are tools that can help. [12]
Figure 3-3 includes a screenshot that shows APISPY32 capturing all calls to strcpy on a target system. We used the APISPY32 tool to capture a series of lstrcpy calls from Microsoft SQL server. Not all calls to strcpy are going to be vulnerable to buffer overflow, but some will.
)
Figure 3-3 APISPY32 can be used to find lstrcpy() calls in the SQL server code. This screenshot shows the results of one query.
APISPY is very easy to set up. You can download the program from www.internals.com. You must make a special file called APISpy32.api and place it in the WINNT or WINDOWS directory. For this example, we use the following configuration file settings:
KERNEL32.DLL:lstrcpy(PSTR, PSTR) KERNEL32.DLL:lstrcpyA(PSTR, PSTR) KERNEL32.DLL:lstrcat(PSTR, PSTR) KERNEL32.DLL:lstrcatA(PSTR, PSTR) WSOCK32.DLL:recv WS2_32.DLL:recv ADVAPI32.DLL:SetSecurityDescriptorDACL(DWORD, DWORD, DWORD, DWORD)
This sets APISPY to look for some function calls that we are interested in. While testing, it is extremely useful to hook potentially vulnerable API calls, as well as any calls that take user input. In between the two comes your reverse engineering task. If you can determine that data from the input side reaches the vulnerable API call, you have found yourself a way in.