DELMAN offers a novel approach to protect deployed language models from jailbreak attacks without compromising overall performance.

• Uses targeted model editing to dynamically respond to detected attacks
• Maintains general task performance while blocking harmful outputs
• Provides post-deployment protection without extensive retraining
• Enables adaptive security that evolves with new attack patterns

This research is critical for secure AI deployment in enterprise settings, addressing the growing concern of adversarial manipulation of language models in production environments.

DELMAN: Dynamic Defense Against Large Language Model Jailbreaking with Model Editing