Mindgard research shows that ChatGPT's image generator can be manipulated to create violent and sexually explicit content without explicit user requests. This raises concerns about the effectiveness of content filters and the implications of training AI models on such imagery.