The How and Why of Container Vulnerability Management

Editor's Notes

• #5: http://www.istockphoto.com/photo/computer-crime-concept-gm516607038-89059287?st=9174601 http://www.verizonenterprise.com/verizon-insights-lab/dbir/2016/ Every year since 2008, Verizon have published a report on the attempted data breaches occurring within their data centers. For 2015, they found close to 90% of them had either a financial or espionage component to them. This report is well worth the read, and there are a few key findings in this report we should all be aware of. Costs of data breaches are heavily skewed towards legal consultation and forensics, and not to the public components of credit monitoring or lawsuits Despite some vulnerabilities having been public for years, there remain vulnerable components in use Some of those components simply may not have a patch forthcoming for a variety of reasons.
• #6: Despite years of organizations spending energy protecting against attacks, it remains up to the attacker to define what’s valuable. Consider the case of ransomware. A police department in the town next to where I live was subjected to a raonsomeware attack. For roughly 500 USD in bitcoin, the attackers would decrypt the booking and evidence records they had just crypto locked. As an attacker, they likely had no knowledge of who they had attacked or what they had locked up. What mattered was the ransom, and that they had a police organization’s files didn’t factor into the equation.
• #7: To recap something we already know – investment is skewed. Attackers target applications while we invest in perimeter defenses. This is a reactive strategy built on an assumption that a gatekeeper will always be better, but fails to understand the reality of what can happen if you’re already on the other side of the wall.
• #8: Let’s take a little bit of time and look at how an attack is created. Potential attackers have a number of tools at their disposal, and use a number of different tactics. In this case, the attacker wishes to create an attack on a given component. In order to be effective, they have two primary models. First they can actively contribute code in a highly active area of the component with an objective of planting a back door of some form. The hope being that their code will fail to be recognized as suspect given how quickly the area of code is evolving. Second they can look for areas of code which are stable, and the longer they’ve bene stable, the better. The reason for this is simple, old code is likely written by someone who isn’t with the project any longer, or perhaps doesn’t recall all assumptions present at the time the code was written. After all, its been long understood that even with the best developers, assumptions change and old code doesn’t keep up. The goal in both cases being to create an attack against the component, so they test, and fail, and iterate against the component until they’re successful or move on. Assuming they’re successful, they create a deployment tool and document the tool for others. Of course, given the publicity received by some recent vulnerabilities, a little PR goes a long way. Now there are responsible researchers who follow a similar workflow, and they legitimately attempt to work with component creators to disclose vulnerabilities. They too will publish results, but are less interested in creating the an attack beyond a proof of concept. http://www.istockphoto.com/photo/person-in-hooded-sweater-using-a-laptop-on-wooden-table-gm464503138-58544934?st=cf78f31 http://www.istockphoto.com/photo/cloud-computing-gm518556682-90104967
• #9: For example, if we have a datacenter filled with servers, and each server has virtual infrastructure which includes virtual networking and security services, we already have multiple layers of perimeter defenses. If we then add in container VMs which run a minimal OS upon which we load up some containers, we have a pretty good model for where some datacenters are today. If we then assume an attacker was able to compromise a container, now what? They’re on the other side of the various perimeters, and can attack from the inside. The big question being precisely how much damage could that attack create, and how would you know? This is the opportunity we have, and the people who are responsible for this are practitioners of a philosophy known as DevOps.
• #11: https://www.cesg.gov.uk/guidance/open-source-software-%E2%80%93-exploring-risk-good-practice-guide-38 If you’re using commercial software, the vendor is responsible for best practice deployment guidance, the notification of any security vulnerabilities and ultimately patches and workarounds for disclosed vulnerabilities. This is part of the deliverable they provide in return for their license fee. If you’re using open source software, that process becomes partly your responsibility. To illustrate the level of information you have to work with, let’s look at a media-wiki maintenance release from December 2015. “various special pages resulted in fata errors” – this clearly is something which needs resolution, but which pages? How do you test? “1.24.6 marks the end of support for 1.24.x” – this is good to know, but I hope it was published elsewhere. “However, 1.24.5 had issues (along with other versions) so it was thought fair to fix them” – This is a good thing, but can we expect this treatment in the future? From the title, we also have a fix for 1.23.x, but what other versions?
• #12: https://sourceware.org/ml/libc-alpha/2016-02/msg00416.html https://sourceware.org/bugzilla/show_bug.cgi?id=18665 https://security.googleblog.com/2016/02/cve-2015-7547-glibc-getaddrinfo-stack.html https://web.nvd.nist.gov/view/vuln/detail?vulnId=CVE-2015-7547 (published via US-CERT) http://cve.mitre.org/cve/cna.html https://openclipart.org/detail/200681/primary-patch
• #13: https://sourceware.org/ml/libc-alpha/2016-02/msg00416.html https://sourceware.org/bugzilla/show_bug.cgi?id=18665 https://security.googleblog.com/2016/02/cve-2015-7547-glibc-getaddrinfo-stack.html https://web.nvd.nist.gov/view/vuln/detail?vulnId=CVE-2015-7547 (published via US-CERT) http://cve.mitre.org/cve/cna.html https://openclipart.org/detail/200681/primary-patch
• #14: https://sourceware.org/ml/libc-alpha/2016-02/msg00416.html https://sourceware.org/bugzilla/show_bug.cgi?id=18665 https://security.googleblog.com/2016/02/cve-2015-7547-glibc-getaddrinfo-stack.html https://web.nvd.nist.gov/view/vuln/detail?vulnId=CVE-2015-7547 (published via US-CERT) http://cve.mitre.org/cve/cna.html https://openclipart.org/detail/200681/primary-patch
• #15: https://sourceware.org/ml/libc-alpha/2016-02/msg00416.html https://sourceware.org/bugzilla/show_bug.cgi?id=18665 https://security.googleblog.com/2016/02/cve-2015-7547-glibc-getaddrinfo-stack.html https://web.nvd.nist.gov/view/vuln/detail?vulnId=CVE-2015-7547 (published via US-CERT) http://cve.mitre.org/cve/cna.html https://openclipart.org/detail/200681/primary-patch https://www.youtube.com/watch?v=hkryI6eapOA
• #18: Datadog, a systems monitoring company, looked at the use of containers from 10,000 of their customers. There are multiple observations in the report, but the most significant is yet more data supporting the assertion containers are now production ready. This doesn’t mean we want to abandon other aspect of the DevOps story, rather it means we need to ensure we’ve a completeness to the story, and that we need to be a serious player in the container space.
• #20: If you assume from the outset your containers will be compromised, what would you do differently? What could you do to make life much harder to mount an attack from a compromised container? The first item everyone should have on their list is to enable the SELinux and AppArmor security modules for the distro you’re using as a base. For SELinux, you specify that it’s enabled on the command line for Docker Engine, and then on the container you specify a security option representing the profile you want to use. Now that you have some security modules in place, you’re in better shape, but you can still make it much harder to mount a proper attack. Consider for example an attacker who recognizes they’re in a container, and assumes that means there probably are multiple containers with the same profile. Armed with that knowledge, one of the easiest ways to make the life of an attacker harder is to randomize the memory load location. That’s where kernel security profiles come into play. Grsecurity, PaX and seccomp enable roles based access and address randomization upon load capabilities. Net of this, where the executable code lives in memory changes, and that makes it harder to know if you’ve created a viable attack or not. Now the Docker people have done a great job over the years in locking down the level of privileged access you get from within a container. In part, this is done using what’s known as kernel capabilities. Kernel capabilities offer a granular level of control over the types of operations you want the kernel to perform. Consistent with the concept of least privilege, you don’t want to ask for more rights than you need. If you find the defaults are providing more access than your application requires, you can pare things back using the –cap-drop option. Of course, it’s entirely possible you might need more rights, but if you find you need to disable priviledged access, or want to set the CAP_SYS_ADMIN flag, beware you’re effectively giving the container the equivalent of root access. Lastly, from a security perspective, you can choose to use a minimal Linux distro such as Atomic or CoreOS, but you still will want to pay attention to the options I’ve just outlined. So now that you’ve limited the access a potential attacker has to system services, you still have to contend with other types of havoc. A perfect example of this is the concept of a noisy neighbor. Most of us have had the experience of having someone in a neighboring space behave in an annoying manner. In the case of computing, the annoying neighbor can be one which consumes excessive memory or processing time. Limiting the scope of interference is very easy. All you need to do is define some CPU shares and memory limits for the container and set them during launch.
• #24: Image source: US Navy 040814-N-5781F-033 Storekeeper 2nd Class Daniel Mina
• #28: Hub is based on a 3 part architecture: Scan Client – scans directories and artifact files creating “code prints” that uniquely but confidentially identify the files & directories contained in them Web Application – This is the main user interface and logic center for Hub. KnowledgeBase – This is a repository for open source component, license, and vulnerability information The code prints recoreded by the scan client are compared to reference code prints in the KnowledgeBase to identify open source components, versions, and origins. No source or binary code is ever uploaded.
• #29: http://www.istockphoto.com/photo/strength-in-unity-gm514713440-88219133?st=af7fa36
• #30: Docker Container Security Scanner https://info.blackducksoftware.com/Security-Scan.html 14 Day Free Trial to Black Duck Hub https://info.blackducksoftware.com/Demo.html Red Hat Atomic Host Integration (Requires Black Duck Hub) atomic scan --scanner blackduck [container]