In very very short conclusion:
Threat modeling should and can be done at almost any point in the development process, but really you should start it early in the process. You should take some "what-could-possibly-go-wrong" or attacker mentality and apply it to your feature.
For the impatient let's have the maybe shortest possible threat modeling guide:
Let's keep it short and simple here by just taking the initial sentence from our Threat Modeling handbook page:
Threat modeling is the process of taking established or new procedures, and then assessing it for potential risks.
This is maybe the most high-level and abstract description of threat modeling, now let's put it to some practical use. We take something, an established or new process, really this might be anything. Like our own assessment-tool, a stand-alone GitLab instance or maybe a new piece of infrastructure (internal link).
For the rest of this HowTo let's call the thing we're doing a threat model for a feature because for GitLab most of the threat modeling which should be done would be on features of the product or features of the infrastructure of our SaaS offering.
What a threat model does, it let's you take apart and decompose the feature you're looking at so that you can, in some more or less formalized way, identify and describe threats towards the feature and its components. This description so far sounds really formal and might not help you to actually come up with a threat model, just take it as a vague baseline of what (identify threats) should be achieved and let's now see how we can get there.
The formal framework that GitLab uses is called PASTA. This guide helps apply the PASTA principals in a lightweight manner to maximize their impact when followed by any GitLab team member.
You might wonder: "When should I start to do some threat modeling on the feature I'm planning to implement?". It's really not obvious but luckily it's never too late and never too early to create a threat model for any given feature which is currently being developed or used. Also, like everything in security, it's more of a process than a fixed, eternal state. A threat model needs to be adopted and refined if the feature changes or even if the feature stays the same but it's being used in a different context or environment. So yes, this is some kind of "extra" work in a sense that a proper threat model just doesn't create itself. But those extra steps will pay off soon enough by yielding more insights about the "what could possibly go wrong" moments which are also known as threats. Even tough it's never too late it will pay off even more if you start the process early and keep the threat model up to date for any additions and changes to the feature. This is also due to the fact that the fixing of insecure design decisions might be rather complex and even disruptive(internal link).
Most threat modeling frameworks rely on some diagrams to be drawn and there's a lot of quite heavy tooling around this. Such tools would take some data flow diagram (DFD) and automatically map certain threats to certain components based on what they are. This can be really useful in some cases, but at GitLab we need some more flexibility as the features we're looking at often would not really fit well into the strict schematics of e.g. STRIDE based threat modeling and thus the output might not yield much meaningful threats.
For diagrams there are almost endless possibilities, for easy and integrated usage within GitLab it'd be recommended to use one of our supported diagram tools. But really to keep it flexible and accurate you can use whatever diagramming tool you're comfortable with. It should just deliver an accurate diagram of the feature. For instance this mermaid diagram of our Kubernetes Agent is completely sufficient as a start. It shows all involved components in their interaction at a level which is still understandable and meaningful.
If we'd keep it too simple like:
It'd still not be wrong but we wouldn't be able to deduce too many detailed threats from this.
One notable component which typically does not come with your typical architecture diagram is a trust boundary. A trust boundary is also not depicted explicitly in the Kubernetes Agent diagram, but we can deduce it. A trust boundary is separating parts in the diagram which have different levels of trust. In the Kubernetes Agent case we can see this clearly, the Kubernetes cluster can be anywhere and be controlled by anyone, it's not really trustworthy. The GitLab controlled components however are controlled by GitLab, therefore very much trusted. So in conclusion we have a trust boundary between those two parts of the diagram. This now is the part where the actual threats come into play. The threats typically manifest at those trust boundaries. A first threat which might come to mind when just looking at this trust boundary:
agentk and kas might be unencryptedThis is a very simple and generic thing to consider, still a legit concern and a first step for a threat model.
For covering more of the threat ground we need to shift a bit the perspective towards a "worst-case" or "what could possibly go wrong" attitude, this generally helps a lot to find more threat scenarios. In the very formalized STRIDE approach the actual threat classes are:
Spoofing
Tampering
Repudiation
Information disclosure
Denial of Service
Elevation of Privilege
Hence the name STRIDE. While we do not use the formal STRIDE framework, we can use these threat classes to get an idea on what should be considered to define the concrete threats in our model.
So spoofing in the Kubernetes Agent example might manifest in one agentk being able to impersonate another. The example from above "The communication between agentk and kas might be unencrypted" would for instance fit in the information disclosure threat class. It's not a matter of thinking about every possible single thing which might go wrong here. One could easily come up with an overwhelming amount of potential and rather obscure threats, but that might be very distracting and we would miss the point of the threat model. We primarily want to aim for those threats which seem likely to occur and those which are really impactful when they occur.
Once the initial steps have been done and we have a first set of well thought out threats we can now use this list during development of our feature to mitigate what we've deemed to be the threats and worst case scenarios.
Here again it's really fruitful to think a bit out of the box: a certain feature might work well on GitLab.com but cause trouble and outages in some self-managed environment which is set up quite differently than our SaaS platform. This can be a threat as well and it's caused due to some environmental change for our feature. This in conclusion means that we should always consider the environment our feature is being used in. That environment might change over time or in different deployment scenarios, and it might not always be friendly and trusted in the first place.
To document our threat model and the results we can keep it simple as well. What's really convenient is an overview table like so:
| Threat | Comments | Test |
|---|---|---|
Unencrypted communication between agentk and kas |
✅ grpc communication is done over TLS encrypted Websockets | |
agentk might be able to impersonate another cluster's agentk |
TBD - review authorization | |
Attacks on gitaly level |
agentk has indirect access to gitaly via kas, this might be abused for injections or IDOR attacks |
TBD - check data flows from agentk towards gitaly |
But really anything is fine as long as you write down the identified threats and potential mitigations as well as the needed next steps. This will help your fellow team members1 to gain a better understanding of the feature in general, in terms of risks it might expose, how to use it and how to NOT use it.
There's also a repository with templates for threat modeling you can use in issues or epics.
All the team members, not only those in the security department which torture you with threat modeling ↩