Drug development AI turned out to be misuse for the development of chemical weapons, and it is possible to identify as many as 40,000 types of 'chemical weapon candidate molecules' in just 6 hours

AI is used in a wide variety of fields, and it is also used to search for drug candidate molecules. By abusing this drug development AI, the results of a shock research that makes it possible to search for 40,000 types of chemical weapon candidate molecules in 6 hours have been announced. It is said that some of the presented candidate molecules were more toxic than the nerve gas ' VX '.

Dual use of artificial-intelligence-powered drug discovery | Nature Machine Intelligence
https://doi.org/10.1038/s42256-022-00465-9

AI suggested 40,000 new possible chemical weapons in just six hours --The Verge
https://www.theverge.com/2022/3/17/22983197/ai-new-possible-chemical-weapons-generative-models-vx

AI used in drug development is designed to search for targeted molecules that reduce toxicity. Fabio Ubina , who studies drug development AI at drug discovery company Collaborations Pharmaceuticals , will adjust the drug development AI ' MegaSyn ' that he developed to search for molecules that increase toxicity rather than molecules that reduce toxicity. Setting. As a result of running this AI for 6 hours, as many as 40,000 chemical weapon candidate molecules were identified.

For the learning of AI used in the experiment, data on multiple chemical weapons including data on nerve gas 'VX' was used. For this reason, many of the identified candidate molecules were similar to VX and existing chemical weapons. In addition, molecules that are not similar to existing chemical weapons and are expected to be more toxic than VX have been identified. From this result, Ubina said, 'The harmless tools involved in the development of useful medicines have been transformed into deadly molecule search tools by reversing their uses.'

MegaSyn also has the ability to virtually sort out identified drug candidate molecules before actually synthesizing them. Therefore, if it is used to search for chemical weapon candidate molecules as in this experiment, it is possible to select highly toxic candidate molecules. In addition, although Mr. Ubina refrains from disclosing the details of the experiment contents to prevent abuse, since free tools such as Python and free drug data sets are used in this experiment, if you are a knowledgeable person It is said that the experiment contents can be reproduced at low cost and in a short time.

Mr. Ubina warns about the situation where it is easy to identify candidate molecules for chemical weapons by using AI, and is calling for discussion of a mechanism to prevent the abuse of AI. In addition, Mr. Ubina has taken measures such as 'formulation of guidelines', 'providing content filters', and 'monitoring of usage' that OpenAI, the developer of the text generation AI ' GPT-3 ', is doing as a reference for AI abuse prevention measures. I have listed it.