Blog
- Part 3
Prompt-injecting my own SRE agent through its logs — and finding the attack landed somewhere I hadn't designed for
A planted 'ignore previous instructions' payload reached the model through a tool I hadn't built the scenario around — and why the system prompt's blanket stance mattered more than the specific attack surface I designed for.
- Part 2
I accidentally ran two copies of my own verification harness against the same cluster — here's how I found out, and what it cost
Two copies of the same eval harness ran against one cluster for twenty minutes without either knowing about the other. How the traces caught it, and how to recover data instead of discarding all of it.
- Part 1
I built a Kubernetes remediation agent that can only open pull requests — here's what the eval runs cost and taught me
kubemend verifies every proposed fix independently instead of trusting the model — and that verification pipeline had its own bugs. Two of them, and what a 30-run eval sweep against a real cluster cost to find.