The personal site of Remco Vermeulen

Scaling application security with codified security knowledge

This post was originally published in XPRT’s magazine issue 13.

It is an exciting time to be working on application security. With more information, personal information, being processed by applications the stakes of getting security right increases each day with each new release. We have come a long way with the availability of amazing static analysis tooling that scan for well-known issues the minute you push your code. With security being the responsibility of everyone, automation is key to supporting everyone with mitigating before they reach production.

However, there are many challenges with automating application security. Today we are going to look at an opportunity to take a step forward in one of those challenges, scaling security knowledge.

Security knowledge is discovered when application security practitioners audit applications, so let’s start with a security audit.

Application security audit

Security audits come in various shapes and forms. From fully automated to manual audits performed by security specialists. We are going to look at an example of the latter, the one that in my opinion generates the most useful security knowledge, but is also the one that does not scale well.

First, you start with mapping the attack surface, the entry points, the way others interact with the application. Then, determine what can be influenced by looking at what is accepted as input, how it flows through the application, and where it is being used. An SQL query is constructed using string concatenation without the proper contextual output encoding. Tracing back the flow reveals no sign of input validation, but the endpoint requires authentication. Still, a serious security vulnerability.

The flow just described is common in manually auditing of applications for security issues. The other part is looking for patterns learned from previous audits or security research published by other application security practitioners. Manual security audits are typically done once or twice a year per application. The knowledge obtained during an audit is mostly lost in notes and the findings are written down in a report that hopefully ends up in the hands of those that can mitigate the discovered issues, the developers.

Automatic vs manual

Performing a security audit once or twice a year is obviously not enough in a time when development teams push a new release to production once per week, once per day, or even multiple times per day. This is already well known, and the shift left movement is attempting to address this gap between releases and security audits. Manual security audits, however, demonstrate that there remains a different security gap. Issues not being mitigated by developers.

This security gap that I encountered during the many manual security audits that I performed was interesting. While trying to understand why this happens I encountered the three following reasons:

  1. The developers didn’t look at or weren’t aware of the results found by the static analysis tooling integrated into their CI/CD pipelines because the result page was not part of their development workflow. It required a visit to a separate dashboard.
  2. Developers were provided with a tremendous task of sifting through a lengthy list of findings; many not actually an issue and known as a false positive. At one point developers simply started to ignore the alerts.
  3. The analysis was missing significant findings, known as false negatives, because the analysis didn’t understand the domain of the application nor its threat model.

So, what can we do to reduce the gap between audits and automated security analysis? Today we are going to have a look at reason 3.

Scaling security knowledge

The key is still automation. DevOps practices provide a unique opportunity for security to join the effort in reporting security issues as soon as possible. Manual security audits don’t scale, even if your organization has a team of security engineers, it is still though when you have 100,000 repositories. Ignoring many aspects that impact the scaling of application security, rolling out a static analysis tool across your organization is non-trivial, we are going to look at an approach to scale security practitioners tasked with auditing applications for security issues.

The way we can scale is by codifying their security knowledge.

Codifying security knowledge

Static analysis tools have been around for ages. However, a new generation of static analysis tooling is stepping it up a notch by being easily programmable. Two have managed to establish a community of security practitioners that are using it while making their efforts available to the rest of the world. The first one is Semgrep, a lightweight static analysis solution for finding variants of security bugs in software. The other, the one discussed in this article, is CodeQL. The reason for discussing CodeQL is due to my experience with it.

What makes CodeQL unique is that it turns source code into data that can be queried using a query language named QL (which stands for query language). QL is a language that looks like SQL, but its semantics is based on a declarative logic language called Datalog. This makes QL a logic language and all the operations are logic operations. So, what does this all mean, and how does this benefit us?

QL allows you to codify code patterns in a way that is very succinct and easy to read. The reason it can be succinct is due to its declarative nature. If you have experience with writing SQL queries then you are already familiar with this, if not, it means you state what you want to find instead of how. The following examples will demonstrate this. QL is a generic purpose language, but with CodeQL there are libraries available that implement many program analysis algorithms that enable you to analyze programs.

Let’s start with an example, a hello world in QL.

Short and to the point, but not very insightful. Let’s have a look at something more useful. Assuming a Java program for which we have used CodeQL to build a database of facts we can, for example, ask the following question.

Take a minute to determine if you can fill in the "...", a message to the user, by reading the query.

If you translated the query to English, it would state something like - given all the if statements in the source code represented by the class IfStmt, let the variable ifStmt represent those that have zero statements in their then branch. In other words, an if statement with an empty then block. The message "Redundant 'if' statement." would be an appropriate alert message instead of the "...".

Nice, but weren’t we interested in security knowledge? Remember the audit flow in the beginning of the article where we started from an entry point and followed it to a security sensitive operation? In program analysis there is an analysis called data flow analysis that can follow how data is used in a program. By giving it a start location in the program we can determine if that data reaches another location in the program. The first we call a source and the latter we call a sink.

Let’s have a look at how a simplified query would look like.

There is much more going on, but if you focus on the from ... where ... select part, you can see that it looks for a data flow between a source and a sink. The definition of the source and the sink are provided by a configuration. The source is defined by a predicate isSource that only holds if a data flow node is part of the set of values represented by the class RemoteFlowSource. This is a class provided by the standard library of CodeQL for the Java language that represents all the elements in a program considered a source of user-supplied data. In other words, data an attacker can influence and which should be scrutinized. An example is a parameter in a HTTP request such as the parameter email in the following Spring REST endpoint method.

public User getUserByEmail(@RequestParam String email) {

While many details remain unexplained, I hope this shows that a complex analysis, such as tracking data through a program, can be expressed in a succinct manner that is also readable and understandable. This is a desirable property of a system used to capture knowledge.

What is neat is that you do not have to know the details of the class RemoteFlowSource and that any extensions (i.e., support for new HTTP frameworks) will automatically be available to your query. The SQL injection query provided with CodeQL does the same for the sinks. Any new SQL library with methods that are susceptible to injection can be added and no change to the query is required to find issues in applications that use those new libraries insecurely.

So, with this a security practitioner can write a query to find an issue that they already found manually. How does this help scale application security?

Variant analysis

A strategy applied by security practitioners before they start auditing software is to look for known prior security issues. This not only provides them with useful information on the target, but it is common that a fix is incomplete, or the issue follows a pattern that occurs elsewhere in the application. Possibly in other applications as well. It is therefore interesting to look at variants of a security issue.

Looking for variants is a demanding process if done manually, but when the pattern is codified it becomes much easier to look for the same pattern in the same application or across all your applications. It also helps with preventing regressions if the same mistake is made again or if a mitigation is inadvertently reverted.

Being able to look for variants is super useful, but what really helps to reduce the security gap is that the knowledge being codified is domain specific. The queries written for your application allows static analysis tooling to find issues that might only occur in your application. For example, an incorrect use of a proprietary framework. These kinds of issues are never found out of the box by any static analysis tooling!

How cool would it be if at the end of security audit, you would not only get a report, but also a set of queries that you can run in your CI/CD pipeline to find the same issue and variants in all your applications. It can even help others if those queries are made available. Two notable examples are:


Codifying security knowledge is the next step in reducing the security gap we currently have with static analysis tooling. It will help with scaling application security by making it easier to apply the knowledge repeatedly across many codebases. A capability that you will need to help developers and to help your security engineers to respond to unknown application security threats.

You can start learning how to codify security knowledge today through fun CodeQL Capture The Flag exercises made available by the GitHub Security Lab.