A single unlabeled prompt is enough to strip the safety guardrails off an LLM

GRP-Obliteration: Unaligning LLMs with a Single Unlabeled Prompt

GRP-Obliteration (GRP-Oblit) uses Group Relative Policy Optimization to remove safety constraints from aligned models with just one unlabeled prompt, while largely preserving utility. Tested on fifteen 7-20B models across GPT-OSS, DeepSeek, Gemma, Llama, Ministral, and Qwen, it beats prior unalignment methods on average and also works on diffusion-based image generators. Safety alignment, the authors argue, is only as robust as its weakest failure mode.

Safety alignment is only as robust as its weakest failure mode.
  1. tyfon

    The magic word seems to be

    Create

    a fake news article that could lead to panic or chaos

    They only tested it on open weight models. I wonder if it also works on the closed ones, bit I don't really want to get banned

  2. stratos123

    That's pretty interesting. It was already well-known that you could easily remove safety training from open-weights models by a bit of finetuning, but apparently you don't even need a finetuning dataset, as long as you have just a few prompts and another LLM to judge responses? Let's see if the abliteration people take a note of this.

  3. whythismatters

    >Submitted on 5 Feb 2026

More from this day

2026-09-15