TokenBreak: Bypassing Text Classification Models Through Token Manipulation