Fingerprinting WAF Rules with Timing Based Side Channel Attacks

0xInfection
Aug 18, 2019 · 12 min read

Hi folks,

Today in this post I am going to be detailing about my recent experiments with web application firewalls (WAFs) focusing on a specific type of side channel attacks, namely, timing. In my opinion, this field hasn’t been studied actively and the results can be more lethal than you’re expecting. This post is already long, so lets get started rightaway. :)

Image for post
Image for post

Side Channel Attacks?

An attack based on information gained from the implementation of a computer system, rather than weaknesses in the implemented algorithm itself.

So basically we are extracting/learning sensitive info which shouldn’t be known to the public using side channel attacks. This is achieved usually due some faulty business logic implementations which are vulnerable to this kind of enumeration attack.

The attack we are going to talk about today is based on the clock. A timing attack focuses on the data movement in and out of the CPU/memory on the hardware running the system or algorithm. Simply by observing time-variations on how long the CPU is taking to analyse data being fed into it, one can enumerate internal sensitive information from the system.

Web Application Firewalls

Why Fingerprint ’em Rules?

Here in this article I have used a common method of rule fingerprinting known as regex-reversing which typically relies on examining every single component of a request to understand which part of the request is causing the block.

Understanding The Setup

  1. Reverse Proxy: The WAF literally sits between the client and the server, intercepting requests. The client connects directly to the WAF and then the WAF passes the query (if normal) to the server. In case of a blocked request, the query never reaches the server.
  2. Server-Resident: This is the setup when a WAF is typically installed on the server it is protecting. This can be further categorized into 2 topologies, the first being that the WAF is installed as a plugin, while the second one is when the WAF is installed as a programming library.
  3. Out-of-Band: In this case, the WAF usually gets a copy of the traffic via a monitoring port on a network device. This mode of implementation limits the WAF’s ability to block a request and can only send TCP-reset packets to interrupt traffic whenever a malicious query is detected.
  4. Cloud-Deployment: This setup comprises of the WAF functioning within the network cloud of the provider. The working is similar to reverse proxy setup, with the exception that every single request to the server has to pass through the network cloud.

In my experiments, I took up two most commonly used implementations of a WAF, namely, the reverse-proxy and the plugin-based server resident setup.

Conventional Methods of WAF Fingerprinting

  • WAF Block Message: Signifying that the WAF has flagged and blocked the request as malicious. Usually the blocked response page or a header defines that the request has been blocked. Sometimes the response status code (403 Forbidden) indicates a blocked request too.
Image for post
Image for post
Some examples of WAF block-pages
  • Web-App Error Message: Signifying that the web-app bugged out upon the request. However the error message gets cloaked by the custom block page of the WAF. In this case, the WAF did not block the request, but only cloaked the error message of the web application to prevent sensitive information disclosure via stack traces, etc.
Image for post
Image for post
Secure Entry WAF cloaking the server’s stack traces
  • Normal Response: Signifying that the request has clearly passed through the WAF to the web server. However, there remains a possibility that the WAF intercepted the request and removed the malicious part of the request before passing it onto the server.

The Main Drawback

Why Timing Attacks?

Idea of the Attack

Principle

Assumption: The only assumption here is that our WAF blocks a request and returns an error message immediately upon detecting a malicious request. Other WAF variants which strip the malicious parts of a request and forward the sanitized request to the server are not being taken into account.

The Approach

Image for post
Image for post
Attack Approach Analysis of our Methodology

Our approach to this problem initially would be to first split the attack into two phases :-

  1. Learning Phase: In this phase we measure and learn what could be the possible response time for a blocked and a passed request for further reference in our attack phase.
  2. Attack Phase: In this phase we perform the actual tests, i.e. the rogue requests are being sent for the final results and further statistical analysis.

Now, comes the computation methodology. In the learning phase, first off, we measure the response times〈T = t, t, … tₙ 〉of n blocked requests and define a “flagging threshold”. This flagging threshold will serve as a future reference when it comes to determining whether a request passed or got blocked. It can be represented as:

Image for post
Image for post

Similarly we define a “passing threshold” by taking n passed requests where the boundary can be defined by taking the minimum value of all the response times of requests which got past the WAF undetected. This threshold can be represented as:

Image for post
Image for post

where in both cases δ signifies that, the actual passing/blocking boundaries can be slightly less than that of the measured values due to possible network latency or noise.

So now theoretically, our flagging and passing boundaries’ timings should serve as a threshold for blocked and passed requests, while practically this is not replicable due to the obvious noise in the network. Thus we take up the mean of both blocking and passing boundaries to get an absolute boundary which can give the actual threshold value about the response time for a blocked as well as a passed request.

Image for post
Image for post

Clearly, any timing measurement for a request with t ≥ t is a passed request while t < tᵦ is a blocked request. However, any request with response time near to the boundary tᵦ may denote a passed request with low noise or a blocked request with high noise. To rule this out, the attacker needs to amplify this attack vector. We’ll discuss more about this, read on.

Performing the Experiment

The Setup

The Learning Phase

Image for post
Image for post

Some examples of passed and blocked/rejected requests are:

Image for post
Image for post
Examples of blocked and passed requests

You may notice that some of the requests in cases of blocked requests crossed the flagging boundary. This is probably due to the internet connection suffering packet losses or network congestion. However, it can be noted clearly that there is no request crossing the absolute boundary on either side.

The Attack Phase

Image for post
Image for post
Comparison how normal and polymorphic payloads look like

Now a series of such requests are sent over to the the server hosting the WAF, and the response times are recorded. To further validate the claims of the experiment, both response status codes as well as the block-pages are recorded. A correctly identified malicious request would result in a 403 status along with the blockpage as response content.

On the reverse-proxy setup, the results were pretty much as expected. There was a minimum time gap of 53.2 ms between a flagged and passed request. The results indicate that in 96.4% of cases, we are able to correctly distinguish between a blocked and a passed request. Thus, we can safely say that the accuracy and reliability of our method for this topology is high enough to assert that any attacker can reach perfect measurement conditions with only a few repetitions.

Image for post
Image for post
Results Visualized for Attack Phase on Reverse Proxy Topology

In the case of the server resident setup (plugin based), I actually did not expect the results to have such distinctive result, however I was surprised to see that the method works quite well as it does in case of reverse-proxy topology. Something worthwhile I noticed in this setup is that the overall timing difference between a blocked and a passed request (58.8 ms) turned out to be greater than the previous topology (53.2 ms). The results obtained for this topology are visualized are as below:

Image for post
Image for post
Results Visualized for Attack Phase on Server Resident Topology

Thus we can safely say that our timing attack clearly distinguished blocked and passed requests in 95%+ of cases. A short summary of my experiment is as below:

Image for post
Image for post
Final Results in a Nutshell

Downsides of the Method

Dealing with It

Amplifying the Attack

1. Choosing a Longer URL Path

2. Denial of Service Attacks

3. Cross Site Rule Fingerprinting

<html>
<body>
<img id="test" style="display: none">
<script>
var test = document.getElementById(’test’);
var start = new Date();
test.onerror = function() {
var end = new Date();
alert("Total time: " + (end - start));
}
test.src = "http://sitename.tld/path?" + parameter + "=" + payload;
</script>
</body>
</html>

In the code shown above, we create an invisible img tag. Just before we duplicate the payload to the URI of the image, we start recording the time. Since the image is invalid, the browser will fire an onerror event handler and the relative function will be executed when the time recording stops and an alert box with the recorded time will be popped up.

There are three main advantages to this method:

  • First, the attackers identity remains concealed. Since multiple users will be sending requests to the server with the WAF, it is virtually impossible to distinguish who is the actual actor behind this.
  • This method absolutely overrules the impact of WAFs blocking IP addresses as a counter security measure.
  • Also, it is notably important to state that this method works reliably only with timing based attacks. Sometimes SOP (Same Origin Policy) might restrict reading page content from other origins. So, in such situations, one may not be able to observe the blockpage and fingerprint the WAF using storage side channels.

Conclusion

So far, thanks for reading and stay tuned! Bye! ⊂(◉‿◉)つ

— 0xInfection

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch

Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore

Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store