Raesene's Ramblings

There’s been a couple of studies recently released by security research companies about exposed Kubernetes clusters on the Internet, and whilst it’s nice to see the security industry focusing a bit more on Kubernetes, some of the analysis misses some of the details of why Kubernetes clusters are exposed to the Internet and what some of the results mean, so I thought it would be a good opportunity to revist this topic, also as there have been some developments in what information can be found via Internet search engines.

There are lots of tools which we can use in the container ecosystem to easily create and test applications, but sometimes the networking they create can get a little complex, making it hard to work with and troubleshoot. I came across a scenario recently (for a workshop in Kubecon) where I needed to access a GUI application deployed in a KinD cluster running in an EC2 instance on AWS, from my laptop. The solution I came up with was to use Tailscale and as it seemed like a nice way to solve the problem, I thought it was worth documenting.

Before I get into this post a quick note that there is no dramatic payoff here, it’s just playing around with something that surprised me in Kubernetes, to understand a bit about what’s going on.

This week there was some research published from Cambridge university called “Trojan Codes”, around the potential risks of RTL unicode characters in source code. Whilst this is very much not a new problem, there have been various pieces of research over the years about the difficulties of handling unicode characters, it seemed like a good cue to look at this kind of issue in the context of Kubernetes. So far I’ve not found any security issues caused by this, but I found a couple of things which could be of interest, so thought I’d write it down, in case it’s useful to anyone.

I was playing around with CRDs today when I wondered “what would happen if I tried to overwrite one of Kubernetes’ core object types? The results weren’t what I expected, so thought it was worth writing it down, in case it comes as a surprise to others :) N.B. There’s an update at the bottom which sheds more light on what happens when you do this

One of the things that comes up a lot with Docker security is that, by default, the authorization model is all or nothing. Essentially any user or process that has access to the Docker socket (or Docker TCP port if it’s configured to listen on a network) can use any of its functions. This means that from a security standpoint essentially any user with Docker access can get root on the host easily.

I had cause to look into unprivileged binding of ports in Docker containers recently, and found a couple of things that were a surprise to me, so taking the time to write them down, in case it’s of use to anyone else.

I’ve always been a fan of spelunking around in Internet data sets to see what interesting things I can find, so when I saw that Censys had a new version of their search site, and that it had classified data about Kubernetes clusters, I had to have a look around :) Before I talk about the results, I will say I was very impressed by their new offering, the search speed is amazing for the amount of data and there’s some really useful search filters, some of which I’m using here.

Container image signing has been a bit of a gap in the security landscape, so I’m always interested in seeing new projects starting up which address it. Docker Content Trust/Notary never really gained traction in v1, and whilst v2 looks very interesting, it’s still in the design phase (AFAIK).

Following on from looking at the escalate verb in Kubernetes RBAC, I thought it would be worth looking at another one of the unusual verbs you can see in Kubernetes RBAC, bind.