Yubin Qu

Fine-tuning as Jailbreaking: A data-centric red teaming framework via logic injection

A red-teaming testing method for fine-tuning-stage defenses that audit the harmful content of poisoned fine-tuning data, exposing a vulnerability of the RFT data supply chain to logic injection and point to the need for fine-tuning-stage defenses that audit the harmful content of poisoned fine-tuning data.