protecting qwen3-8b from gcg based persona jailbreaks by steering with a linear direction

Researchers developed a method to protect the Qwen3-8B language model from GCG-based persona jailbreaks by applying a linear steering technique that guides the model's behavior away from undesired outputs.