Exploring dual-use risks of large language models for new psychoactive substances: a severity-based vulnerability analysis

Recent advances in large language models (LLMs) have profoundly transformed scientific research in chemistry, enabling literature mining, the prediction of physicochemical and biological properties, and even the generation of novel molecular structures. However, such tools also raise concerns regarding malicious use, particularly in the context of the clandestine synthesis of new psychoactive substances (NPSs). In this study, we evaluate simulated scenarios of interaction with LLM-based chatbots to explore potential vulnerabilities in the provision of information that could facilitate illicit synthesis. Inspired by recent methodologies that apply LLMs to chemistry, we designed conceptual experiments involving different types of queries, ranging from neutral consultations to explicit attempts to obtain hazardous instructions. Our findings indicate that, although the models often refuse to provide direct answers, risks remain associated with the release of indirect information, bibliographic references, and technical details that could be exploited indiscriminately. The discussion emphasizes ethical and regulatory implications, proposing the need for red teaming strategies, risk-oriented databases, and cross-sectoral cooperation to mitigate potential misuse. This study contributes to the emerging debate on digital security in chemical sciences and outlines future directions for the responsible use of artificial intelligence.