This research introduces PiCo, a novel jailbreaking framework that exploits visual modality vulnerabilities and code training data characteristics to bypass security measures in Multimodal Large Language Models (MLLMs).

• Employs a tier-by-tier approach to systematically overcome defense mechanisms
• Demonstrates significant security gaps in current MLLMs
• Exposes the tension between capabilities and safety in multimodal systems
• Highlights the need for more robust defense strategies against emerging attack vectors

This research matters for security professionals because it reveals critical vulnerabilities in increasingly deployed multimodal AI systems, providing insights for developing more comprehensive safeguards against sophisticated attacks.

PiCo: Jailbreaking Multimodal Large Language Models via Pictorial Code Contextualization