Sora Inappropriate and Harmful Content Creation Easily Bypassed Through Simple Prompt Engineering

In this paper, we demonstrate how Sora, a text-to-video generative model, can be prompted to generate inappropriate and potentially harmful content despite the presence of safety mechanisms. Through a systematic application of simple prompt engineering techniques, we were able to bypass content restrictions by rephrasing, obfuscating, or embedding cues in seemingly innocuous ways. Our findings reveal critical shortcomings in Sora's content moderation layer and highlight broader implications for the deployment of generative models at scale. We discuss the methodology used, present representative examples of successful bypasses, and analyze the ease with which these manipulations can be reproduced. Finally, we reflect on the ethical responsibility of developers and propose potential mitigation strategies to reinforce safety without compromising creativity.