Colluding LoRA: A Compositional Vulnerability in LLM Safety Alignment

Sihao Ding

We show that safety alignment in modular LLMs can exhibit a compositional vulnerability: adapters that appear benign and plausibly functional in isolation can, when linearly compos…