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

Area of Science:

• Artificial intelligence in chemical sciences
• Digital security in chemistry
• Responsible AI development

Background:

• Large language models (LLMs) offer advancements in chemical research, including literature mining and molecular design.
• However, LLMs also present risks for malicious applications, such as facilitating the clandestine synthesis of new psychoactive substances (NPSs).

Purpose of the Study:

• To evaluate the potential vulnerabilities of LLM-based chatbots in providing information that could aid illicit chemical synthesis.
• To explore risks associated with simulated interactions and different query types.

Main Methods:

• Designed conceptual experiments simulating user interactions with LLM-based chatbots.
• Employed varied query strategies, from neutral inquiries to direct requests for hazardous synthesis information.
• Analyzed model responses for the potential release of exploitable data.

Main Results:

• LLM chatbots frequently refused direct requests for illicit synthesis information.
• Significant risks were identified through the indirect release of bibliographic references, technical details, and related data.
• Vulnerabilities exist even when models attempt to block harmful content.

Conclusions:

• The study highlights the need for robust security measures against the misuse of AI in chemistry.
• Recommendations include implementing red teaming strategies, developing risk-oriented databases, and fostering cross-sectoral collaboration.
• Responsible AI development is crucial to mitigate potential security threats in chemical sciences.