Reading Between the Pixels: Linking Text-Image Embedding Alignment to Typographic Attack Success on Vision-Language Models

This project is an academic study of a known vulnerability in VLMs (typographic attacks, such as the 'Granny's house' prompt injection or simple OCR-based exploits). While the paper provides valuable analytical insights by linking attack success to embedding alignment, it is currently a reference implementation with no stars and minimal forks, indicating it has not yet gained broad community traction. The defensibility is low because it is an analysis of a flaw rather than a proprietary solution or a tool with a moat; the findings can be easily replicated or integrated into safety training by larger labs. Frontier risk is high because labs like OpenAI and Google are already aggressively working on 'Visual Jailbreaking' and OCR safety as part of their core safety benchmarks. The displacement horizon is short (6 months) because the next generation of native multimodal models (e.g., GPT-4o, Gemini 1.5) are specifically designed to handle these visual-textual discrepancies better than the models likely tested in this study.