A read-only audit of my own observability stack, run like an engagement. Almost every serious thing it found was inside a defence I had written hours earlier, and the worst one was inside the proofs I use to show the defences work.
I pentested my own AI hub and shipped the method, not the map
