Researchers from Shandong University have introduced "Safe-Control," a plug-in-style security protection solution designed to safeguard text-to-image generative models from producing unsafe content. As text-to-image models rapidly advance, they are capable of generating high-quality images—yet they also pose significant risks. When given malicious or inappropriate text prompts, these models can produce content involving violence, pornography, hate speech, and other harmful material, which may mislead users and negatively impact society. To address this growing concern, Ph.D. student Xiangtao Meng and his team have developed Safe-Control, a modular security patch that can be seamlessly integrated into existing text-to-image models. The solution operates in real time during image generation, effectively suppressing the creation of unsafe content without compromising the model’s output quality. Unlike traditional safety mechanisms, Safe-Control demonstrates strong transferability across different models and remains effective against a wide range of adversarial text prompts. Extensive experiments have shown that Safe-Control significantly reduces the likelihood of unsafe image generation across multiple leading text-to-image models. Reviewers praised the work for tackling a critical issue in generative AI and for proposing a novel, plug-in-based approach. The method’s high degree of generality and practical applicability make it a promising tool for real-world deployment. As text-to-image models continue to improve in quality and versatility, their use is expanding into sensitive domains such as advertising, education, entertainment, virtual social platforms, and even finance and healthcare. However, this growth also intensifies the need for robust safeguards to ensure generated content remains compliant and safe. Safe-Control offers a reliable defense mechanism for these applications, helping maintain ethical and legal standards in AI-generated content. The research journey began with identifying the core challenge: how to enhance model safety without degrading performance. The team explored various strategies, balancing the need for strong content filtering with the preservation of high-quality image generation. This required careful design to avoid any negative impact on the model’s core capabilities. After numerous iterations and extensive testing, the team settled on a plug-in architecture that injects safety control signals into the generation process. This approach allows the security layer to be added without modifying the original model’s structure. The team faced significant technical hurdles, including designing effective training datasets, defining appropriate safety criteria, and integrating control signals efficiently. Each step was validated through rigorous experimentation and parameter tuning. The final version of Safe-Control proved highly effective, consistently reducing unsafe outputs across diverse models while maintaining strong generalization. The success of the project reflects not only technical innovation but also the power of collaboration and creative problem-solving. The research was guided by Professors Zheng Li and Shanqing Guo, who provided critical insights and inspiration. A pivotal moment came during a discussion when Prof. Li proposed drawing an analogy to software “patching” in operating systems—a concept that helped break traditional design constraints and led to the development of the plug-in framework. Prof. Guo contributed practical solutions that helped overcome key implementation challenges. Looking ahead, Meng plans to build a community around Safe-Control to expand its use cases and enhance its resilience against emerging attack patterns. He also aims to deploy the solution in real-world industrial settings, helping organizations manage the risks associated with generative AI. Since joining the School of Cyber Security at Shandong University, Meng has focused on the security of large-scale AI models, earning recognition through multiple high-impact publications at top-tier conferences such as IEEE S&P and CCS. His work has established a strong foundation for future research in AI safety and model defense. Currently, he is a Ph.D. candidate dedicated to advancing the security, reliability, and compliance of generative AI systems.