27th United Nations Science and Technology Conference

The 27th United Nations Science and Technology Conference on Information (STI) held in Switzerland saw the World Digital Technology Academy (WDTA) release two international standards for the safe testing of generative artificial intelligence (AI) applications and large language model (LLM) safety testing.

The two generative AI safety standards were written by researchers from institutions such as Ant Financial Group, OpenAI, Meta, NVIDIA, Amazon, Microsoft, the University of California, Georgetown University, universities, and institutions, including Tencent, Baidu, OPPO, iFlytek, CloudWay, Leeloo, Google, Hangzhou Huabei Technology, KPCB, Reco AI, Exabeam, NovNet, BreachQuest, Anthropic, Kainos, Precize ai, Private AI, the University of Chicago, Carnegie Mellon University, the Hong Kong University of Science and Technology, the University of California, Berkeley, WDTA, CSA, ISACA, SIG, the Center for Inclusive Change, the University of Chicago, OWASP Foundation, and others Organization audit,  a total of 36 companies, universities and institutions around the world participated in the compilation.

WDTA is an international non-governmental organization registered in Geneva and follows the guidance framework of the United Nations, dedicated to promoting digital technology globally and promoting international cooperation. The AI STR (security, trustworthiness, and responsibility) initiative is WDTA’s core proposal, aimed at ensuring the safety, trustworthiness, and responsibility of AI systems

It is reported that this is the first time in the world that global standards have been released specifically for generative AI and large language models. It fills the gap in the field of security testing for large language models and generative AI applications, providing a unified testing framework for the industry. They can provide clear testing requirements and methods for AI enterprises, helping to improve the security of AI systems, reduce potential risks, promote the responsible development of AI technology, and enhance public trust.

The first standard is the “Generative AI Application Security Testing Standard,” led by WDTA and jointly developed by Ant Group, OpenAI, Meta, NVIDIA, and Microsoft, among others. This standard provides a framework for testing and verifying the security of generative AI applications, especially those built using large language models (LLMs). It defines the scope of testing and verification for each layer of the AI application architecture, including the selection of base models, embedding and vector databases, RAG or retrieval-enhanced generation, and the security of AI application runtime, among others.

The second standard is the “Large Language Model Security Testing Method,” led by Ant Group. This standard provides a comprehensive, rigorous, and practical structural solution for the security assessment of large models themselves. It proposes a classification of security risks for large language models, a method for classifying and grading attacks, and a testing method. It also provides the first classification standard for four different levels of attack intensity. It systematically provides evaluation indicators, ability grading, data set construction requirements, and testing procedures for the resistance of large language models to attacks.

Through this standard, it can effectively address the complexity inherent in large language models, comprehensively test and verify the resistance of large language models to different types of adversarial attack techniques, including L1 random attacks, L2 blind box attacks, L3 black box attacks, and L4 white box attacks. This enables developers and organizations to identify and mitigate potential vulnerabilities and ultimately improve the security and reliability of artificial intelligence systems built using large language models.