False Positives Everywhere
As a web developer, I've learned to ignore vulnerability scan results and that's a big problem.
Most of the results aren't relevant. Sometimes I find it easy to figure out what's safe to ignore from the findings. Other times I find myself scouring documentation, source code, and blog posts only to discover the "RCE" npm audit told me I had doesn't matter.
Sifting through the list of false positives, looking for the slim possibility of an actually impactful vulnerability isn't how any of us want to spend our time. I always assumed that there was no better way. The infrastructure just didn't exist to eliminate that tedious job.
At LunaSec, we've been taking on this type of work for a few select companies, and taking notes. When diving in deeper to this problem we realized something a little mind-blowing.
Scanners could do a lot better
Vulnerabilities are usually tied to a library or a package, at least the ones that end up in an official database and show up in your scan report. Typically, the CPEs (packages) that are vulnerable are written right into the CVE (the official report).
Most scanners consume this database. If they find you have a vulnerable package, they notify you. Full stop.
We've reviewed thousands of these notifications. We wrote down the reasons results could be ignored, and in the end realized that the vast majority of these false positives fell into one of a few categories. Many of them could be eliminated by relatively simple code.
Common false positive scenarios
Here are some common reasons why a reported "vulnerable package" might not be exploitable:
Transitive dependency eliminates vulnerabilities
Most dependencies installed in our projects aren't the direct dependencies that we've asked for at the top level. Most are the dependencies of those dependencies, and so on. They're deeper in the tree. Since this is where most dependencies are, this is where most vulnerabilities tend to occur as well.
Often the relationship of a vulnerable package in connection to its parent, helps us to identify unrealistic vulnerabilities. Maybe the parent package doesn't call the vulnerable code or pass in the un-sanitized input needed to trigger it. Other times, you identify it as moot due to its relationship in the dependency tree. Wow, a Regexp Denial of Service vulnerability found in a sub-dependency of my testing library? I don't care! Why are you showing me this?!
Take this vulnerability in one of React's utilities from just a few days ago:
The critical information most people need is right there in the CVE description for you to read, but that's not machine-readable.
Language version or operating system not vulnerable
Often, especially with languages like Java, a vulnerability will only matter for some versions of the language. We have seen plenty of vulnerabilities that are only relevant to apps running on Java 9 or higher. With somewhere around two-thirds of projects still using Java 8, we can stop notifying of those vulnerabilities two-thirds of the time!
Vulnerable function not called
Sometimes you'll have a vulnerable library installed, (or one of your dependencies has it installed) but the vulnerable function isn't being used. This one is a little harder to search for, but thanks to tools like Semgrep, it's easier than you'd think to validate that you aren't using these vulnerable functions in your code.
The description of a CVE will sometimes sometimes contain potentially problematic functions that you can watch out for. With SemGrep, those usages could be codified and scanned for.
Tracking these common scenarios
We need to come up with a machine-readable way to capture the above information, before we can scan for it.
The National Vulnerability Database (NVD) was established 17 years ago, and at the time it was groundbreaking. But, it wasn't built for use with automated scanners. Now we need a more structured standard to catch up with the much larger quantity of Open Source packages. We also need much more of a focus on machine-readable data.
At LunaSec, we more-or-less owned the early discourse around Log4Shell and some other high-profile vulnerabilities through our blog posts, and in doing so we saw just how eager people were to submit details and provide information. A structured format would have made those blog posts much easier for us to maintain. When a major vulnerability hits, the abundance of blog posts, POCs exploits and vulnerable repositories, discussions on Twitter, are all proof that the effort and interest is there.
For now, we're thinking of simple Yaml files in a GitHub repository to capture the highest-severity vulnerabilities. Here's our take at codifying Log4Shell as structured information. We call them "guides" because they capture more information about a vulnerability than a simple report.
Building a graph of ALL dependencies
That's right, all of them. We think.
We're in the process of cloning every NPM and Maven package. With this we will be able to graph out the relationships between dependencies. Within those relationships we can begin to add the context needed to eliminate false positives. We could, for instance, flag all denial-of-service type vulnerabilities that occur in sub-dependencies of popular testing frameworks, such as Jest.
It also will open up other supply-chain security approaches, such as hosting a repository mirror ( like Artifactory), detecting / preempting dependency hijacking, and license scanning. The sky is the limit.
Building a more contextually-aware scanner
All of these scenarios have one thing in common, and it's that they require more context. Being able to scan someone's full source-code is already a step up from just scanning a package inventory like a manifest or SBOM. The best case scenario, though, is scanning an entire system, and that means scanning docker containers.
When you scan a container you get the full system context including what language version is installed, what system packages are present, what the Linux distribution is, and so on. Collecting container information and using this to supplement code scanning will help remove false positives.
At the moment, we've written a CLI that can scan and upload data about built containers (or any file or folder) from a CI job, but eventually we would like to host a container repository (think DockerHub) where people can push their containers to be scanned.
What we have done so far
Our dependency scanning tool, LunaTrace, is currently in beta and you can try out. It has GitHub integration, CLI support to integrate with your CI job, and you can even drag and drop a file, folder, or zipped container right into the web app to see instant results.
The groundwork is there to start building these false-positive elimination strategies. Our plan is to start with the lowest hanging fruit and work our way up to a quiet tool that only asks for a human when one is truly needed. We see plenty of work to be done.
Perhaps we are being overly optimistic, and we can maybe only eliminate half or two thirds of the false positives. Even so, that seems like a tool we'd prefer to use.