Group 1 - The AI field is transitioning from the era of large models to the era of intelligent agents, which brings security challenges such as overreach and excessive delegation [1] - The "2025 Terminal Intelligent Agent Security" white paper was jointly released by Shanghai AI Laboratory, CAICT, Ant Group, and IIFAA Alliance, providing a comprehensive risk assessment guide for terminal intelligent agents [1][2] - Intelligent agents are rapidly penetrating various terminal devices like smartphones, glasses, headphones, and car systems, redefining interaction methods across multiple industries including life, industrial, medical, and education [1] Group 2 - The white paper outlines three major protective paths: single intelligent agent security, multi-agent trusted interconnection, and AI terminal security, aiming to serve as a comprehensive and targeted security guideline [2] - The white paper introduces a terminal intelligent agent security system supported by a technical ecosystem, detailing security technologies for single agents and multi-agent interactions [2] - Over 70% of intelligent agent practitioners express concerns about issues like AI hallucinations, erroneous decisions, and data leaks, with more than half indicating their companies lack a designated security officer for intelligent agents [3] Group 3 - Ant Group's "Ant Tianjian" has announced an upgrade to its large model security solution, adding intelligent agent security assessment tools with a risk judgment accuracy rate exceeding 96% [3]

Group 1 - The forum on July 28 focused on the advancements in large model technology, including intelligent agents, multimodal applications, and open-source initiatives, highlighting the global trends and directions in large model innovation [1] - The China Academy of Information and Communications Technology (CAICT) aims to promote high-quality development of the large model industry through systematic research, infrastructure development, and international governance participation [1] - Shanghai's Xuhui District is implementing strategies to build a national AI hub, creating an innovation ecosystem and enhancing policy support and talent cultivation to accelerate the establishment of a large model industry ecosystem [1] Group 2 - A "Large Model Open Source Security Initiative" was launched by CAICT in collaboration with various organizations to address security risks associated with the large-scale application of open-source models, promoting a safe and collaborative ecosystem [2] - The release of the report "Terminal Intelligent Body Security (2025)" marks the first comprehensive study on the security of terminal intelligent agents, proposing a multi-layered security protection system to foster innovation and safety in intelligent applications [2] Group 3 - The forum was guided by the World Artificial Intelligence Conference Organizing Committee and supported by various local government and industry organizations, establishing itself as a premier event in the large model industry [3]

Core Insights - AI glasses are viewed as the next generation of smart terminals following smartphones, with significant interest from major tech companies [1] - The "2025 Security of Terminal Intelligent Agents" white paper highlights privacy risks associated with smart glasses, emphasizing the need for security measures [1][3] - The evolution of AI agents is transforming various industries, leading to more proactive and human-like interactions [2][5] Group 1: Security Risks - The white paper identifies multiple layers of security risks for terminal intelligent agents, including device, data, model, cloud collaboration, interface, and business layer risks [7] - Specific risks include user identity, application identity, malicious prompt engineering, permission abuse, behavior inducement, and uncontrolled content generation [7] - The emergence of prompt injection attacks poses a significant threat to terminal intelligent agents, necessitating enhanced security measures [7] Group 2: Technological Advancements - AI agents are transitioning from traditional smart devices to more integrated, end-side intelligent agents that redefine user interaction [2][5] - Innovations such as gesture recognition and 3D semantic understanding are enhancing user experience in devices like Meta Quest 4 Pro and Huawei Vision Glass 2025 [2] - The shift from passive to proactive user services is changing the interaction dynamics across various sectors, including automotive and healthcare [5][8] Group 3: Collaborative Security Measures - The white paper proposes three main protective paths: single intelligent agent security, trusted interconnection of multiple agents, and AI terminal security [8] - Key dimensions for building a trusted collaborative environment include trusted connection, authentication, intent, and authorization [8][10] - The establishment of the IIFAA Trusted Interconnection Working Group aims to address data calling, identity verification, and permission boundary issues across multiple devices and accounts [9][10]

Core Insights - The report emphasizes the transition of intelligent agents towards more endpoints, showcasing enhanced generalization and adaptability across various applications, evolving from single-task functionalities to complex scenarios such as programming assistance, social interaction, and economic governance [1][3] - The emergence of lightweight intelligent agents, capable of running directly on everyday devices like smartphones and wearables, is becoming a mainstream trend, particularly in consumer-facing applications where user concerns about privacy, response speed, and personalized experiences are increasing [3][4] - The concept of multi-agent systems is highlighted as a development trend, where multiple intelligent agents collaborate to complete complex tasks, necessitating a systematic ecological architecture for efficient communication and cooperation among agents [4][5] Industry Trends - The Shanghai World Artificial Intelligence Conference (WAIC) has brought attention to the accelerated application of endpoint intelligent agents, while also highlighting the new security challenges posed by extensive connectivity and complex architectures [1][3] - The report calls for enhanced research on endpoint intelligent agent security and the establishment of a comprehensive, multi-layered security protection system to ensure the healthy development of the large model industry and intelligent agent applications [1][5] - The need for a collaborative security ecosystem is emphasized, advocating for a framework that includes data sharing, capability collaboration, and standard recognition to address systemic security challenges in the context of cross-domain cooperation and evolving threats [5][6]