It’s long been proven that LLMs can produce exact copies of their training data. Not every piece of training data, of course, but enough where spitting out exact replicas of proprietary code is a legitimate concern. Not to mention that copies don’t need to be exact in order to be a problem.