LATIDIA · Ciberseguridad

Ataque de DPO verdaderamente benigno de pocos disparos para LLM de Jailbreaking

arXiv: 2605.10998v2Announce Type: replace Resumen: Las API de ajuste hacen que los LLM de frontera sean fáciles de personalizar, pero también pueden debilitar la alineación de seguridad durante el ajuste. Mientras que el trabajo previo muestra que benigno supervisado fino-

La noticia

arXiv:2605.10998v2 Announce Type: replace Abstract: Fine-tuning APIs make frontier LLMs easy to customize, but they can also weaken safety alignment during fine-tuning. While prior work shows that benign supervised fine-tuning (SFT) can reduce refusal behavior, deployed fine-tuning pipelines increasingly support preference-based objectives, whose safety risks remain less understood. We show that Direct Preference Optimization (DPO) introduces a stronger and harder-to-audit failure mode. We propose a truly benign DPO attack using only 10 harmless preference pairs, the minimum data scale accepted by OpenAI's fine-tuning service. Each pair contains a benign prompt, a normal helpful answer as the preferred response, and a refusal as the dispreferred response. Unlike prior benign fine-tuning attacks,