Open-Weight LLM Fine-Tuning Defenses are Susceptible to Simple Attacks
Current open-weight LLM safeguards are brittle because they only focus on preventing harmful behavior after fine-tuning, ignoring pre-existing harmful knowledge. The study shows simple gradient-free attacks like Abliteration and Prefilling can bypass these defenses with high success rates. The proposed solution, Abliteration-Resistant Tuning (ART), improves safety by training the model to resist these specific low-cost attacks.