If the result is that nobody runs security testing on AI models for fear of being prosecuted, have we really solved the problem?