This research reveals how vertical text formatting can be manipulated to bypass content moderation systems powered by large language models.

• Vertically aligned text significantly reduces LLMs' classification accuracy
• This vulnerability affects various models including GPT-3.5/4 and Claude
• The technique could be exploited to evade harmful content detection
• Researchers propose defensive techniques to strengthen LLM security

For security teams, this highlights the need for more robust content filtering systems that can handle text formatting variations. Understanding these vulnerabilities is crucial for developing more resilient AI safety measures.

Vulnerability of LLMs to Vertically Aligned Text Manipulations