Immune introduces an innovative inference-time alignment approach that significantly improves safety in multimodal LLMs vulnerable to jailbreak attacks.

• Creates a safety filter that detects and neutralizes harmful requests without disrupting normal operations
• Addresses the critical gap between training-time alignment and real-world attack scenarios
• Demonstrates effectiveness against sophisticated visual-based jailbreak attempts
• Offers a practical solution that can be implemented without retraining existing models

This research is vital for security professionals as multimodal AI systems become more widespread, providing a defense mechanism that works alongside existing safety measures to protect against evolving attack patterns.

Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment