Researchers have proposed a new defense mechanism called DiSCO to prevent text-to-image generative models from producing harmful content. The system operates entirely at the prompt level and doesn't require any changes to the model itself. It works by optimizing prompts through contrastive scoring over safe and unsafe image pools generated by the target model, with iterative adaptive feedback until safe content is produced. DiSCO has been shown to enhance the safety of both undefended and defended models on a benchmark under multiple red-teaming attacks. --- Why it matters: This matters because text-to-image generative models can generate harmful content such as violence or nudity, and existing defenses have limitations in scaling to proprietary models. DiSCO's ability to operate as a plug-and-play module makes it a promising solution for addressing these safety concerns. Source: https://arxiv.org/abs/2608.17067