When harmless words cause harm: the new threat of concept trigger words to large language model safety

The researchers propose a model-agnostic jailbreak method named MICM, which uses concept trigger phrases to manipulate the implicit value stance of model outputs, inducing inappropriate content without triggering conventional safety filters. The method is based on concept morphology theory, encoding predefined phrases into a fixed prompt template to guide model outputs toward a specific value stance. The authors evaluated it on five advanced large language models including GPT-4o, DeepSeek-R1 and Qwen3-8B, reporting that MICM consistently outperforms state-of-the-art jailbreak techniques, achieving a high success rate with very few refusals. The authors state that this finding reveals that the safety mechanisms of commercial large language models remain vulnerable to covert manipulation of underlying value alignment, and that existing alignment strategies have blind spots against attacks exploiting models' abstract generalization capabilities.