In the previous sections of the course we have shown how input checking is important and introduced some tools for Fuzzing the input. As a reminder, dumb fuzzing was based mainly on random inputs on a black box and smart fuzzing relied on test harnesses on the target application. What happens inside an application when the input is given to it? If one follows the target application with a debugger one can easily see that data is being copied and modified here and there inside the program. One can even say that data is always moving (most of the time anyway). One should also remember that the data may be code and it may be accidentally interpreted if there's a vulnerability in a parser.
The key to understanding how a complex piece of software works is drawing a picture. You might have heard about a saying that goes "a picture is worth a thousand words". Humans cannot remember many things at a time and most systems have legions of parts and some of them are produced by a third party. It helps if the developers and the security analysts can sit down and draw while discussing about the system and what it is doing. Resolving questions like "how" and "exactly what" are very effective while one draws on whiteboard and others comment.
What to draw and how
The diagrams presented in this section are more conceptual than technical as there is no need for all the details and the form of it is free. For example, class diagrams are not useful with threat analysis as they do not relate to processes, i.e. run-time encapsulation. Although some information from the classes is useful, such as serialization of data, but it can be expressed easily by writing it to the diagram.
Data Flow Diagram
Data Flow Diagrams are graphical representations of the data flow through the system. It is used to get an overview of the system but can be used to visualize the data processing. From the diagram it should be visible what kind of information is going in and out of the system.
Message Sequence Chart
What if there are more communicating entities and/or more complex protocols with more messages. The Message Sequence Chart is an interaction diagram which is used to illustrate communication system components and their message interchange flows. In MSC all the entities have their own vertical line and between them the messages are drawn as horizontal arrows between the entities. In MSC time runs from top to bottom.
What to include in the diagram
We just got off from saying that there is no need to be fully formal and there is no need to draw unnecessary details into the diagram. Hmm, so what to draw into the diagram? What kind of information is then needed?
There might be available architectural images of the whole system and they are tempting starting points to use in the analysis but actually they are not. These images might come from design team and possibly they are even for marketing purposes and should not be used. In real-life what the system exactly does is not always clear.
When starting with the architectural threat analysis the people around the table rarely agree on what everything in the system does. Also the fact that the developers use abstractions that are helping them to work efficiently but the abstractions might mask what is really happening in the software or give wrong assumptions on the workings of the system. Sometimes the design plan has been changed and the changes are not translated into the system itself or the system contains undocumented debugging features, etc. The best way is to start from the empty whiteboard, draw the overall architecture, and refine it in steps. These steps include many questions of "how" and "what exactly" and they are repeated multiple times. How many times depends on the situation but the point is to arrive in a place where all around the table agree that the security analysis level is meaningful.
So what to draw then? Keeping in mind that the result should explain the system in sufficient detail. This may differ case by case but at least the following should be found from the diagram.
Let's start from the way how the attacker interacts with the target system. The data needs to flow to and from the attacker to the system. This is called an interface, which is a place in the system that crosses a boundary. A security boundary is a barrier between processing blocks that is enforced externally, for example, OS controls that a process cannot read another processes' memory.
The first letter in DFD stands for data, but what do we want to know about it? It is good to know three things about the data: where did it come from, where is it going, and what it is. First, where did the data come from? If the data came from outside it can be attacker controlled. Second, where is the data going to? This helps in figuring out the places in the system that may have potential attacker controllable parts. Some good follow-up questions to this are how is the data transported and where is it going to be stored. It makes a huge difference that the access rights for important files are correct or that the database users are correctly setup. Third, what the data is? Depending on the data it might need more or even more security services. The data may contain information that needs to be kept safe. Information such as social security numbers or other data that needs to be kept private or otherwise sensitive data such as credit card numbers. Normally data is the content in the system but also cryptographic keys, certificates, and configuration files of the system should be considered as data.
The DFDs are full of blocks or sometimes bubbles (also known as bubble charts). These blocks handle the data processing and they have to be reduced to the point in which they contain a single decision on the data. This basically means that you are going as deep as needed until all ambiguity is removed from the block. Important things to draw with the blocks is information such as the implementation language. Also where did the code or block come from might be of interest as applications written today reuse existing code and are based on frameworks. From the diagram it should be easy to understand what framework and what version was used and what are all the static and dynamic libraries. Moreover, plug-ins and extensions should not be forgotten. What not to reduce? When you find a processing block over which you are not developing or have no control over otherwise. For example, the browser can be left undrawn if the system is purely server side application. However, the data flow to the browser has to be drawn.
The underlying technologies are one point of interest in DFDs. Virtual machines, load balancers and such should also be drawn in. Basically anything that handles the data in a way or another.
The DFDs are full of boxes but what are they and what do they contain? In data flow analysis, the boxes are boundaries that act as barriers between processing blocks. Some of the typical boundaries:
- Machine boundaries
- Containing boundaries
Barriers are usually externally enforced. Security boundaries are nested meaning that the physical machine is the first boundary and for example the virtual machines running on that machine are boundaries and processes running on those virtual machines are boundaries, etc. As mentioned before, in OSs processes cannot read other processes memory (this is a boundary only if the OS or the filesystem really enforces access control). Which would mean that threads do not have their own security boundaries as they share memory with their parents. Somewhat similarly same owner's processes can access the same files in the filesystem (not a boundary if looked from the filesystem point-of-view). Machines and virtual machines are also boundaries as processes do not move between machines or virtual machines.
Containing boundaries are boundaries that do not come from the system but are usually created on purpose. Containing boundaries are isolated from the hosts execution environment in some way. For example a containing boundary can be a chroot jail, a Mandatory Access Control system (MAC), a sandbox or an operating-system-level virtualization system.
Java VM and seccomp are examples of sandboxes. A quarantine sandbox looks for malicious files in the confined area and determines if it is something bad. This sandbox is more like a brief period in which the files are scanned and then released. Operating-system-level virtualization can also be thought as a sandbox. For example, Docker creates the illusion of containers being their own machines by taking advantage of chroots, resource namespaces and resource usage limitations.
AppArmor and SELinux are examples of MAC. With them you can harden the processes running on the same OS and enforce detailed system checks whether the processes have the correct rights to access resources. One might ask why to use this kind of hardening in addition to the built-in OS access control but the reason is quite simple. When operating normally the OS keeps the application (their processes) at bay but if an attacker finds a way to exploit a weakness in the process and tries to access resources it is not meant to read then the MAC will not allow the operation. For example, the access control rules can specify which files, devices, and system calls are allowed and which not.
If an attacker is able to bypass a boundary, it usually is safe to assume that the attacker is in control of all the access rights and what happens inside that boundary. While accomplishing this may not be straightforward, it makes sense to assume the worst has happened. Moreover, this is a good incentive to keep your system up-to-date and hardened. For example if the attacker somehow gains access to a JBOSS Application Server running on the machine the attacker gains control over all the applications running on it or if the attacker gets the control of the kernel of that server the attacker gains control of the whole machine (including the virtual machines on the physical machine).
All the data flows have to end somewhere. Termination is an important concept as always when the data flow is parsed or acted on is a possible place for a attack. These termination points are the places where the target system is exposed to external influence. For this reason it is important to understand where each data flow terminates.
Protocols are one part of interest in the data flow analysis. Typical data flows are using networking protocols such as, IP, UDP, TCP, SSL, TLS, HTTP (Ethernet is not mentioned as it is not end-to-end protocol instead it terminates switch by switch and connects only devices in close proximity). These protocols are stacked on top of each other and all of them have their own task in the networking stack (Please see the OSI model). IP is the responsible one to getting the packet routed from the source to the target, Next layer is the transport layer which handles things such as flow control, reliability, multiplexing with either UDP or TCP depending on what kind of traffic characteristics are needed. Protocols such as TLS are used to provide confidentiality and integrity on top of TCP (DTLS on top of UDP). On top of that protocols such as HTTP and CoAP are used to move web page requests and responses. On top of there may still be application level protocols that are using all of the above to transport their protocol messages.
But where do these protocols terminate or do they just pass through? Data passing through if it is passed through without touching it (i.e. no parsing, filtering, etc.). These points are not part of the attack surface. Most of the IP routers and such belong to these devices. Anything that looks into the traffic, be it a proxy or a load balancer terminates the flow of that layer and makes it a possible point of attack. From a black box it is hard to determine if it is passing through the data or is it actively "tampered" with. Sometimes for some boxes the "looking into what it is doing" is impossible and these should be treated as untrustworthy termination and new traffic points.
Taint analysis is a form of information flow analysis that is used to track untrusted input through the system (the term originates from perl). A flow is an operation or set of operations that operate on value x to derive another value y. If the x here comes from untrustworthy source it is tainted and it is given a tag or a label. This is done to all user data. These tags allow the tracking the influence of the objects through the target application. Tainting also propagates and is transitive as when a tainted object is used to derive yet another value that value is then also tainted.
What can be done with tainting? We can track user data and see if non-trusted data can reach locations that it should not, i.e. which boundaries does the tainted data end up in. For example if non-trusted data reaches privileged location it is possible that it may cause buffer overflows, XSS, etc. Taint analysis is nice as it can detect that something is wrong even with unknown attacks. In taint analysis all statements are checked if there is tainted objects included and if there is the execution is halted.
The taint analysis phase of threat analysis follows all the data flows and all layers of flows and can identify the components that process the data. Taint analysis is one of the easiest ways to understand which components in the system are the ones exposed most to the attacker generated data. The output of the taint analysis is a good list of candidates for fuzz testing (discussed in earlier part). Other analysis can also be performed on the components in the list.
Data lifetime is very important to think about when implementing software. The lifetime depends on the fact on who allocates the memory for that piece of data and where does it end up. The memory can be allocated from different places, from Kernel, system applications, and from user applications. This means different things for the lifetime of the data and where it ends up. The data may end up in places such as the kernel, any application, system swap, or hibernation storage.
The problem of minimizing the lifetime of a piece of data is hard. One can follow two different approaches the coarse-grained and fine-grained control for propagation. First, in coarse-grained control one cannot determine when the data is for example swapped out of the memory or have little control in core dump situations (written to a log file). Against the swapping one can try to protect by encrypting the swap with dm-crypt or similar if possible or by explicitly locking the data to memory which is possible for example in C by using mmap and mlock. If the core dump is setup incorrectly the contents may be readable to unprivileged users and there is a risk that the dump contains sensitive information. Also in situations where the dump is sent to another machine, for example with netdump, care must be taken that the dump is not sent with "NETDUMPKEYEXCHANGE" variable set into none. Second, the fine-grained control relies on programmers. Depending on the language it may not always be possible but, for example in C this would mean removing memsets. More information of this and related things can be found from the Secure Coding Guidelines for Java SE.
During this course, we have taken an introductory tour to securing (web) software. In the course project, some of this knowledge is taken into use. Stay tuned.
Remember to check your points from the ball on the bottom-right corner of the material!