What Is the New AI Agent Safety Testing Standard and Why It Matters

The World Digital Academy (WDTA) officially released the AI STR series' new standard, the AI Intelligent Agent Operation Safety Testing Standard , jointly led by Ant Group, Tsinghua University, China Telecom and co‑authored with more than twenty institutions worldwide, marking the first global single‑agent safety testing standard.

The standard tackles behavioral risks that arise when agents cross language barriers by linking five key components—input/output, large models, Retrieval‑Augmented Generation (RAG), memory, and tools—to the operating environment, establishing a comprehensive risk‑analysis framework. It also defines agent risk types and introduces novel testing methods such as model detection, network communication analysis, and tool fuzz testing, filling a gap in agent safety standards.

The standard was announced during the AI for Good Global Summit, co‑hosted by the United Nations Institute for Social Development (UNRISD) and WDTA.

2025 is being called the “Year of the Agent,” as AI agents gain deep‑thinking, autonomous planning, decision‑making, and execution capabilities, shifting from “I ask, AI answers” to “I ask, AI does.”

Nevertheless, security concerns are significant: over 70% of agent practitioners worry about hallucinations, erroneous decisions, and data leaks, and more than half report that their organizations lack a dedicated AI agent safety officer.

WDTA Executive Director Li Yuhang warned of a “Colin‑Griich dilemma,” where rapid technology adoption can exponentially increase governance costs, and emphasized that the AI STR standards aim to embed ethics and responsibility throughout the AI lifecycle.

Ant Group’s Large‑Model Data Security Director Yang Xiaofang noted that while AI applications are accelerating, the absence of a unified full‑link safety testing standard makes risks hard to quantify; the single‑agent standard serves as the minimal viable unit for AI governance, with multi‑agent governance envisioned as the future “skyscraper” architecture.

Unlike other international standards, AI STR not only identifies and grades risks but also provides end‑to‑end management solutions—from data governance to model deployment—accompanied by concrete testing tools and certification processes to enhance safety and trustworthiness.

The standard offers a reliable safety benchmark for the global AI agent ecosystem, with early assessments already applied in finance and healthcare sectors.

Previously, WDTA released three AI STR standards: the Generative AI Application Safety Testing Standard, the Large Language Model Safety Testing Method, and the Large‑Model Supply‑Chain Security Requirements, developed with experts from OpenAI, Ant Group, iFlytek, Google, Microsoft, Nvidia, Baidu, Tencent, and others.