The 27th Session of United Nations Commission on Science and Technology for Development (CSTD) was held on April 15-19. The side event “Shaping the Future of AI” on April 16 was hosted by the World Digital Technology Academy (WDTA), an NGO that promotes digital technologies and global cooperation, where breakthrough results including two standards were released.

The two standards for the security of large language models (LLMs), “Generative AI Application Security Testing and Validation Standard” and “Large Language Model Security Testing Method”, were the first of their kind published by WDTA, marking the new benchmark for LLM security evaluation and testing around the globe.

Multiple experts and scholars from OpenAI, Ant Group, iFLYTEK, Google, Microsoft, NVIDIA, Baidu, Tencent and other enterprises have devoted to the development of the two standards. And Ant Group has contributed to the development of the standard for LLM security testing method with leading efforts.
The testing method standard provides a comprehensive and rigorous structural scheme with high operability. It delineates the security risk classification and methods of classification, grading and testing of attacks in LLMs. Also, it puts forward the criterion distinguishing attacks of four different intensities, together with strict assessment indicators and testing procedures, to deal with the inherent complexity of LLMs and fully test their ability to defend against hostile attacks. Thus, developers and organizations can identify and remedy potential vulnerabilities, and improve the security and reliability of LLM-based AI systems.

Big tech companies should play a key role in the secure and responsible development of AI, promote best practices by their resources, expertise, and influence, and establish an ecosystem prioritizing security, privacy, and morality, said Wang Weiqiang, General Manager of Machine Intelligence Department at Ant Group.