University of Modena Proves AI Safety Geometry Transfers Across Every Image Model
Researchers show safety guardrails learned in one LLM can be transferred to image and video generators without ever touching unsafe training data
Takeaways
- First cross-model safety steering framework: Safety vectors learned in a source LLM (Llama, Mistral, Qwen) transfer to image/video generators with no unsafe target data needed.
- Strong ASR reductions: On Flux1-Schnell, attack success rate drops from 0.307 to 0.038; on Wan2.2 video, from ~0.75 to ~0.07.
- Matches or beats native baselines: Transferred directions often rival or exceed safety directions trained directly on the target model with unsafe data.
- Three alignment methods tested: SVD (strongest safety), ridge regression, and MLP — each offering a different safety-vs-quality tradeoff.
- Multi-vector extension: Category-specific safety directions (sexual, violence, hate, etc.) enable selective, fine-grained control per deployment context.
- Project page live; no code released yet — methodology is reproducible from the paper.
Every time a new image or video model drops, safety teams face the same exhausting problem: build the guardrails from scratch, again. Existing approaches require model-specific retraining or carefully curated unsafe datasets for each new architecture. A new paper from researchers at the University of Modena and Amazon Prime Video asks a more fundamental question: do different models encode "safety" in the same geometric place? If so, you could learn a safety fix once and ship it everywhere.
The answer, it turns out, is largely yes. The team published a paper introducing the first framework for cross-model safety steering, demonstrating that a safety direction extracted from a source language model can be transferred to a completely different text-to-image or text-to-video generator, with no unsafe data required on the target side.
The problem with today's safety stack
Existing approaches remain largely model-specific, requiring retraining or tailored interventions for each new architecture. This is expensive and creates a dangerous lag: every new model release means another safety engineering cycle. The deeper issue is that most safety work treats each model as a black box, ignoring the possibility that the internal geometry of "safe" vs. "unsafe" might be shared structure across models.
The theoretical foundation here comes from two converging lines of research. First, early studies showed that targeted activation perturbations can reliably steer model behavior, and further analysis supports the hypothesis that semantic features exhibit linear structure in representation space. Second, the Platonic Representation Hypothesis , the idea that sufficiently capable models trained on similar data tend to converge toward shared internal representations of high-level concepts , suggests that safety geometry might not be model-specific at all.
How cross-model safety steering works
The core question is whether a safety direction learned in a source model remains meaningful when mapped into the latent space of a different target model, even when source and target differ in architecture, training data, pretraining paradigm, or tokenizer, and without access to unsafe data in the target.

- Estimate a safety direction in the source LLM. The team feeds paired prompts into a source LLM (e.g., Llama 3.1-8B) , one safe, one unsafe, differing only in a single policy-sensitive attribute (e.g., "a man lying on a sidewalk" vs. "a bloodied corpse on a sidewalk"). They compute the mean difference between the hidden-state activations of safe and unsafe prompts. This average displacement vector points from "unsafe" toward "safe" in the source model's representation space.