49款国内外AI大模型安全测试结果：我国大多处于中游水平

Core Insights - The 2025 Global Digital Economy Conference highlighted the current state of AI model security in China, indicating that it is mostly at a mid-level with an attack success rate of approximately 5%-6%, suggesting room for improvement [1][3]. Group 1: AI Security Challenges - Experts noted that AI is amplifying security threats, making them more scalable and targeted, which exacerbates the asymmetry in cyber defense [3]. - There is a significant gap between China's AI security algorithms and international standards, particularly in originality, model openness, and dataset quality [3]. - A new security assessment platform for AI models revealed that advancements in AI capabilities have not corresponded with improvements in security risk prevention [3]. Group 2: Importance of Systematic Defense - Experts emphasized the need for a comprehensive network security system, moving away from isolated defense strategies to a unified defense framework [4]. - The successful zero-incident network protection during the 2022 Winter Olympics was cited as an example of the effectiveness of a joint defense system [4]. - Recommendations for establishing a joint defense system include ensuring operators fulfill their security responsibilities and transforming individual service purchases into a public service under a unified defense framework [4]. Group 3: Industry Collaboration and Governance - The importance of collaborative principles in industry chain integration was stressed, advocating for deeper integration of technology and international standards [5]. - A shift from individual product selection to a systematic approach in security construction was proposed, addressing the fragmentation in security procurement among large enterprises [6]. - Suggestions for enhancing security operations include leveraging AI for threat intelligence and establishing efficient feedback loops for security operations [6].