This research identifies and addresses fundamental weaknesses in current safety fine-tuning approaches for Vision-Language Models (VLMs).

• Reveals a safety reasoning gap where models fail to properly analyze visual content in safety-critical contexts
• Introduces a novel Multi-Image Safety dataset specifically designed for training VLMs on safety scenarios
• Proposes an effective Safety Reasoning Fine-tuning approach that improves model safety without compromising helpfulness
• Demonstrates significant reduction in attack success rates while maintaining model utility

For security professionals, this research offers crucial insights into protecting VLM deployments against manipulation while preserving functionality in safety-critical applications.

Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models