Evidence · DiggingBeagle record
Few-Shot Truly Benign DPO Attack for Jailbreaking LLMs
May 2026 preprint on refusal degradation from Direct Preference Optimization with innocuous training pairs; historical study, not a newly patched issue or live attack campaign.
- Published
- May 9, 2026
- Publisher
- Yoon et al.
- Source role
- primary disclosure
Evidence record
May 2026 preprint on refusal degradation from Direct Preference Optimization with innocuous training pairs; historical study, not a newly patched issue or live attack campaign.
Claim-level citations (1)
- supportsBenign preference pairs can reduce safety refusals during model tuning: The authors report using ten harmless preference pairs in controlled API fine-tuning experiments and measuring evaluated attack success from 54.80 percent to 81.73 percent depending on the tested model.
Abstract: ten benign preference pairs