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

Cybersecurity researchers have discovered a novel attack technique called TokenBreak that can be used to bypass a large language model's LLM safety and content moderation guardrails with just a single character change. "The TokenBreak attack targets a text classification model's tokenization...