This research exposes critical security flaws in DeepSeek models by demonstrating how embedding manipulation can force AI to hallucinate specific visual content that isn't present.

• Identifies specific vulnerabilities in vision-language integration mechanisms
• Implements a systematic optimization approach to induce controlled hallucinations
• Demonstrates successful attacks against a leading open-source multimodal LLM
• Highlights urgent need for embedding-level security measures

This work reveals concerning security implications for responsible AI deployment in professional settings where visual content accuracy is essential, such as medical imaging, content moderation, and surveillance systems.

DeepSeek on a Trip: Inducing Targeted Visual Hallucinations via Representation Vulnerabilities