Researchers have discovered a significant vulnerability in Large Vision-Language Models (LVLMs) using automatically generated flowcharts that can bypass safety mechanisms.

• FC-Attack leverages flowcharts with partially harmful content to induce models to generate unsafe responses
• Successfully tested across multiple leading LVLMs including GPT-4V and Claude
• The attack achieves up to 92.2% success rate on certain models
• Existing defense mechanisms prove insufficient against this novel attack strategy

This research highlights critical security implications for AI deployment in enterprise environments, demonstrating how visual content processing remains a vulnerable attack surface even in safety-aligned systems.

FC-Attack: Jailbreaking Large Vision-Language Models via Auto-Generated Flowcharts