Research
Jailbreak to Protect: Buffering and Reinforcing via Temporary Jailbreaking for Safe Fine-Tuning in Large Language Models
Publication date unknown · Discovered openreview.net
Page published
Publication date unknown · First observed: 12 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
Harmful fine-tuning attacks can strip safety alignment from deployed LLMs offered via Fine-tuning-as-a-Service, and this research presents a gradient-level defense that defenders of such platforms can use to preserve refusal behavior.
The ICML 2026 paper "Jailbreak to Protect" proposes Buffer-and-Reinforce, a safe fine-tuning framework that temporarily jailbreaks an LLM during user fine-tuning to saturate safety-degrading gradients, then reinforces refusal behavior afterward. Using BufferLoRA and ReinforceLoRA with QR decomposition-based merging, it defends against harmful fine-tuning attacks in Fine-tuning-as-a-Service while preserving user-task performance.