Introduction

As the capabilities of cutting-edge AI models continue to advance by leaps and bounds, AI security concerns have expanded from social governance issues—such as bias, hallucinations, and misinformation—to encompass cybersecurity, the protection of critical infrastructure, and even national security. Recently, Anthropic’s Mythos model has drawn significant international attention for its enhanced capabilities in vulnerability discovery, attack chain simulation, and long-term autonomous tasks, highlighting concerns about AI-enhanced cyber capabilities. What technical barriers have these cutting-edge models overcome in terms of their cyber capabilities? How should pre-deployment security testing identify genuinely dangerous capabilities while avoiding the creation of new avenues for abuse? Do open-source models amplify risks, or do they provide the necessary conditions for independent auditing and defensive innovation? Against the backdrop of intensifying AI competition between China and the United States, can the two sides establish limited yet pragmatic cooperation? In this episode of SinoAI Insights, we will take the security controversies sparked by Mythos as a starting point to conduct an in-depth analysis of the risks, defenses, and potential for China-U.S. cooperation regarding cutting-edge AI cyber capabilities.

Host

·XIAO Qian, Deputy Director, Center for International Security and Strategy, Tsinghua University

Guests

·XU Wei, Vice Dean, Professor, Institute for Interdisciplinary Information Sciences, Tsinghua University

·FU Hongyu, Director, AI Governance Center and the Data Economy Center, Alibaba Research Institute

Shownotes

02:26 The Mythos Event as A Rehearsal for Future Frontier AI Safety Risks

05:17 Future AI Governance Focuses on "Use Control"

09:10 From Innovation to Control? How Mythos Reshapes the Landscape of US AI Policy

12:46 Chinese Perspectives on Frontier Model Evaluation, Cyber-Capability Thresholds, and AI Agent Risks

14:39 The Paradox of Model Safety: Testing May Create Even More Dangerous Models

18:05 Full-Lifecycle Governance Frameworks and China's Governance Approach

23:07 Using "Isolated Environments" to Ensure Safety and Confidentiality During Testing

25:49 Pros and Cons of the US Granting Limited Access to Mythos

27:48 Three Games Within the US Pre-Deployment Safety Review System

34:48 Risks and Value of Open-Weight Models

46:26 Prospects for US-China Coordination on Model Evaluation

Team

Executive Producer：XIAO Qian

Producers：WANG Yexu、LIU Yuan

Podcast Editor：XU Ruijia

Content Assistant：CAI Leyao

Graphics Designer：ZHANG Shuoning