Model merging is a technique for combining several task-specific LLMs into a single model, allowing developers to assemble specialised capabilities without retraining. The new paper notes that these task-specific models are often sourced from unverified public platforms, which exposes a critical supply-chain attack surface: any malicious behavior can potentially enter the final model through one of its components.

The paper, titled RogueMerge, proposes robust and unified attacks against LLM model merging. It frames the merging process itself as the target, arguing that the reliance on unverified public model sources makes the pipeline vulnerable to attack.

The significance is that as model merging becomes more common, the supply chain for LLMs inherits the same trust problems that plague traditional software supply chains. The paper is a warning that the convenience of merging public models comes with security risks that need to be addressed.