New TokenBreak Attack Bypasses AI Moderation with Single-Character Text Changes

A new attack technique called TokenBreak can bypass large language model (LLM) safety guardrails with a single character change. By altering input words slightly, such as changing "instructions" to "finstructions," attackers can manipulate text classification models while preserving meaning for humans and LLMs. This vulnerability affects models using BPE or WordPiece tokenization but not Unigram. Researchers recommend using Unigram tokenizers, training models with bypass examples, and monitoring for suspicious patterns to defend against TokenBreak attacks.