if ya put it like that, it sounds more like a fundamental failing

The thing is that perfect anonymization implies that the dataset would be useless since by definition it would contain no information. If you can begin correlating data points with enough outside information you will be able to extract at least a shadow of the original information.