You're evaluating a new security tool, maybe a fancy WAF or some cloud-native runtime protection. The sales deck is slick, full of logos and "AI-powered" buzzwords. They show you a dashboard with pretty graphs.
Ask them for their threat model. The one they used to design the thing. Not a generic OWASP top ten list they slapped in an appendix.
Watch them squirm. 9 times out of 10, they can't produce one. Or they hand you a vague document that basically says "bad people try to do bad things."
If a vendor doesn't have a clear, written threat model stating:
* What they are *actually* protecting against.
* What they are *explicitly not* protecting against.
* The assumed capabilities of the attacker.
* The specific system components they're securing.
Then they're selling snake oil. They built a solution in search of a problem. Their detection rules and mitigations are guesses. You'll spend months tuning out false positives and will get owned by an attack vector they never considered.
It's the first gate. Fail it, end the call. Save your time for vendors who do the boring, foundational work.
Example of what I want to see, at minimum:
```yaml
Threat Model: "Container Runtime Security Agent"
Assumptions:
- Attacker has initial foothold on a worker node (via app exploit).
- Attacker cannot modify the hosted Kubernetes control plane.
- Agent runs as root in host PID namespace.
Protected Assets:
- Host system binaries and kernel integrity.
- Uncompromised container workloads on the same node.
Explicitly Out of Scope:
- Supply chain attacks on container images pre-deployment.
- Network layer attacks (e.g., service-to-service encryption).
- Control plane API server compromise.
Adversary Capabilities:
- Can spawn arbitrary processes in a compromised container.
- Can attempt container breakout via kernel exploits.
- Can attempt to disable userland security monitoring.
Mitigations Implemented (mapped to threats):
- syscall blocking -> container breakout.
- immutable filesystem detection -> host binary modification.
- agent self-protection -> monitoring disablement.
```
No map, no journey. Walk away.
If it ain't broke, don't 'upgrade' it.
Spot on about threat models. Where I get cynical is when they *do* produce a beautiful 50-page threat model doc. Perfect STRIDE breakdown, lovely diagrams.
It's pure theater. They outsourced it to a consultancy *after* the product shipped, to check a box for enterprise procurement. The actual engineering team never saw it. The mitigations don't map to the model at all.
—EB
Absolutely. The threat model ask is a fantastic filter, but honestly, I find it only works on the vendors who are naive enough to be embarrassed.
The polished ones? They've gamified it. The threat model is just another asset in the sales kit, decoupled from the engineering sprint cycles. You're right to be cynical.
My favorite move now is to take their beautiful document, pick a random mitigation claim, and ask them to show me the specific commit or feature flag that implements it. The silence is even more telling than the squirm. It's all compliance theater, not security engineering.
Trust but verify.