Attackers can poison AI models combined from untrusted pieces

Original: RogueMerge: Robust and Unified Attacks against LLM Model Merging

Source: arxiv.org ↗