Whitepaper · Threat model
Governing AI agents in Kubernetes
Your agents hold real credentials. Nothing in most clusters checks whether an action is part of the task they were given.
- 9 attack paths, every one from a documented incident
- The coverage matrix that shows where you're exposed
- 8 questions to ask of your own cluster
32 pagesFrom the creators of Kyverno
Get the threat model
Free PDF · Instant download · 32 pages
Loading form…
The form couldn’t load. Please contact us instead.
Your copy is ready
A copy is on its way to your inbox too.
Download the PDF (opens in a new tab)Want us to run the 8 questions on your cluster? →Inside: the coverage matrix
Nine attack paths. Four control points. Some paths only one of them can stop.
4 of 9attack paths have exactly one control point standing in the way. Do you run it?
Unlock the matrixAttack pathBuildDeployRequestExecution
- prevents
- only control
- partial
- blind
July 2026 · The Hugging Face intrusion
AI agents escaped a sandbox and took over production Kubernetes. No attacker involved.
Sources: OpenAI and Qualys, Aug 2026
- <13hrs
- One pod to host-level access across clusters
- 11days
- From escape to first alert
- 1credential
- Cluster-admin on two clusters at once
What's inside
- Agent threat modelSection 3
- Where controls stopSection 4
- 9 attack pathsSection 5
- Coverage matrixSection 6
- Keeping policy currentSection 7
- 8 cluster questionsSection 9
Get the whitepaperAlready read it? Request an agent inventory or see AIControls