GRP-Obliteration Technique Revealed: Removing LLM Safety Alignment with a Single Prompt

Researchers including Mark Russinovich proposed the GRP-Obliteration (GRP-Oblit) method, demonstrating that safety constraints in safety-aligned LLMs can be effectively removed using only a single unlabeled prompt.

Core Mechanism and Performance

• Utilizes Group Relative Policy Optimization (GRPO) to directly remove safety constraints from the target model.
• Achieves stronger unalignment on average compared to existing state-of-the-art techniques while preserving most of the model's utility.
• Generalizable not only to language models but also to diffusion-based image generation systems.

Evaluation Scope and Results

• Evaluated on 15 models with 7-20B parameters (including GPT-OSS, DeepSeek, Gemma, Llama, Ministral, Qwen, etc.).
• Included both instruct and reasoning models, as well as dense and MoE architectures.
• Confirmed that the method overcomes the limitations of existing methodologies through 6 utility benchmarks and 5 safety benchmarks.