Study shows that the safety guardrails of the models crumble after simple fine-tuning

Internal safety mechanisms initially survive innocuous fine-tuning, but LoRA-type protections eventually give way on a larger scale.

An article published at the beginning of September looks at the fragility of the security devices of the major language models in the face of fine-tuning, including when it is carried out in a framework presented as harmless. The authors show that the internal representations related to security remain at first preserved, before deteriorating significantly when the extent of the tuning increases.

Notably, protections that are supposed to limit this risk, such as LoRA or ASAM techniques, are also proving to be less and less effective as fine-tuning intensifies. The study illustrates the difficulty of ensuring that a model that is aligned remains so after a subsequent adaptation phase, even if carried out in good faith.