In October 2026, Tianjiao JIANG, Associate Professor at the Fudan Development Institute and Research Fellow at the Center for Global AI Innovative Governance, was interviewed by People’s Daily Overseas Edition on recent “boundary-crossing” incidents involving frontier AI models worldwide and on global AI governance trends. JIANG noted that the successive incidents of AI models breaking out of security-isolated environments and accessing external systems beyond their authorization mark a new stage in the development of large AI models: the “model loss of control” risk, once a science-fiction scenario, is becoming a real challenge for the industry. Traditional safety designs targeting a single model can hardly cope with the complex security risks brought by multi-agent collaboration, and the incidents also highlight the difficulty of defining responsibility for unintended AI consequences. He identified four core categories of AI security risks currently facing the world: technology misuse, model loss of control, social disruption, and geopolitical and military risks. Although the United Nations, the G20 and other multilateral mechanisms have begun relevant governance exploration, the world still lacks unified consensus principles and coordination mechanisms; the harmonization of AI safety regulation rules and standards remains at a preliminary stage, and the major AI powers differ markedly in regulatory paths—US policy exhibits a swing effect, the EU takes a strict legislative path, and China adheres to balanced governance that gives equal weight to development and security. JIANG argued that addressing AI security risks urgently requires globally coordinated action: promote the establishment of globally recognized third-party independent evaluation institutions for frontier models and mutual recognition mechanisms, so as to form unified international safety standards; establish standardized international mechanisms for reporting and disclosing safety incidents, clarifying the reporting obligations of enterprises and institutions; and improve AI model classification and grading standards, moving beyond judging risk merely by parameter scale and focusing on substantive risk features such as autonomous cyberattacks and high-risk multi-agent operations, so as to draw clear and precise safety red lines for AI development.

Translated by FDDI

Full text in Chinese available at:

https://fddi.fudan.edu.cn/1f/1e/c18965a794398/page.htm