GRP-Obliteration: Unaligning LLMs with a single unlabeled prompt
GRP-Obliteration: Unaligning LLMs with a Single Unlabeled Prompt
This paper introduces GRP-Oblit, a method that uses GRPO to strip safety alignment from LLMs. A single unlabeled prompt reliably breaks safety guardrails while largely preserving utility. Evaluated on 15 models (7-20B) across six families—GPT-OSS, distilled DeepSeek, Gemma, Llama, Ministral, Qwen—it beats existing SOTA on average across five safety benchmarks. The attack also works on diffusion-based image generators. Authors are from Microsoft, led by Mark Russinovich. The abstract doesn't name the specific safety benchmarks or quantify the utility drop; I'd wait for replication before drawing strong conclusions.
Why it matters: A Microsoft team proposes GRP-Obliteration, which uses GRPO to strip a model's safety alignment with a single unlabeled prompt, and compares it against existing methods across fifteen 7B to 20B models.