This research exposes how typographic elements in images can be manipulated to inject unauthorized prompts into cross-modality generation models, creating significant security risks.

• Demonstrates successful prompt injection in both Large Vision Language Models and Image-to-Image generation models
• Reveals that text elements in images can override intended model behavior
• Identifies vulnerabilities that persist despite existing safety mechanisms
• Proposes potential defense strategies to mitigate these threats

This work is crucial for security teams as these vulnerabilities could enable malicious actors to bypass AI safety guardrails, potentially generating harmful content or extracting sensitive information through seemingly innocent images.

Exploring Typographic Visual Prompts Injection Threats in Cross-Modality Generation Models