This study reveals that safety-aligned behaviors in LLMs are controlled by multiple linear directions in activation space, not just a single dimension as previously thought.

• Safety mechanisms use multi-dimensional representation to refuse harmful queries
• Researchers identified specific vulnerabilities in safety alignment by analyzing these dimensions
• Understanding these dimensions provides insights into how jailbreak attempts can bypass safety guardrails
• The multi-dimensional approach offers a more complete framework for analyzing LLM security measures

For security teams, this research provides a deeper mechanistic understanding of how safety alignment works, enabling more robust defenses against manipulation and more effective safety fine-tuning techniques.

The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Safety Analysis