Dearbhadh Retrospective

Over the course of this year I’ve been running some research on how well LLM models understand Kubernetes security and how well agents can secure and attack Kubernetes clusters. It’s been very interesting to see how this has developed over the year so I thought it was worth writing up. Dearbhadh I started this project in January 2026 as part of research for a talk at Kubecon EU 2026. The initial idea when I submitted the talk in late 2025 was to focus on knowledge based tasks, but with the release of Opus 4.5 and the resultant increased interest in agents that could directly take actions as directed by models, I decided to also cover some agentic tasks. The name is part of a theme I’ve used for other projects which is to use Scots Gaelic names, as you can be fairly sure no-one else will have a project by the same name! In this case the name means “proof” on the idea that I was looking for proof of how well models handled Kubernetes security. To get access to the models I was going to…