Preventing Verbatim Memorization in Language Models Gives a False Sense of Privacy