Researchers introduce cross-model safety steering for generative AI

A team of computer scientists has introduced the first framework for cross-model safety steering in generative AI, addressing the challenge of model-specific safety controls. The method, tested on text-to-image and text-to-video generators, begins by learning a portable 'safety direction' from a source large language model (LLM) using paired safe and unsafe prompts. This direction is then transferred to a target generator through a lightweight alignment procedure fitted exclusively on benign (safe) data. Crucially, the framework never accesses unsafe data on the target side, isolating whether safety representation geometries are shared across models. Beyond a single global direction, the researchers also developed a multi-vector extension that captures category-specific safety behaviors, enabling more selective control over content like violence or nudity.

Results across diverse model pairs show that transferred safety directions achieve comparable reductions in Attack Success Rate (ASR) and maintain similar CLIP-Score and FID trade-offs to directions learned natively on the target model using unsafe data. This indicates that safety improvements do not degrade generation quality. The findings suggest a modular view of safety: safety-relevant behaviors are not purely model-local but persist across architectures as latent directions. This opens the path toward lightweight, reusable safety mechanisms that eliminate the need for target-side unsafe data, potentially simplifying deployment of safer generative models in production.