A new preprint on arXiv describes a backdoor attack that does not require poisoning training data with artificial trigger patterns. Traditional attacks insert visible or invisible triggers into samples, causing misclassification only when those triggers appear. This approach instead embeds the backdoor into the natural structure of clean data, making it far harder to detect.

The attack is called "support-persistent" because it exploits the support vectors or similar decision-boundary elements that persist when a model is reused or fine-tuned. The authors argue that models shared for reuse can carry hidden behaviors even if their training data looks perfectly clean. This poses a risk for open-source model distribution and transfer learning pipelines.

Since this is a single preprint abstract, the full methodology and experimental details are not yet available. The claim is notable, but it has not been peer-reviewed or independently replicated. Still, it suggests that clean data alone does not guarantee a model is free of hidden backdoors, especially when models are repurposed. Further scrutiny will be needed to assess the practical threat and potential defenses.