允中 发自 凹非寺 量子位 | 公众号 QbitAI OpenClaw、Moltbook彻底火了。 2026年一开年，AI突然变了——从原来只会写诗、画图的聊天机器人，进化成了"行动式智能体"，能接管操作系统，能自主调用API，能替你 发邮件、管财务。 但当AI真正拥有了"手脚"，恐慌也随之而来。 如果它误删了核心数据？如果它被一句话诱导"黑化"了？ 传统"打补丁"的安全思维，在自主决策的智能体面前会瞬间失效。 智能体安全 ，正成为AI下半场最具挑战的赛道，也是智能体经济规模化落地的必经之路。 近日，江苏 通付盾 提出了一套极具前瞻性的智能体安全框架。 他们认为：AI产业化的前提，正从"能力优先"转向 "信任优先" 。 目前，这套逻辑已在通付盾最新的 "大群空间"（LegionSpace）多智能体协同平台 中落地。 在万亿级智能体经济爆发前夜，我们该如何构建一个既强大又听话的"数字大脑"？一起往下看。 智能体安全体系建设 人工智能已从技术突破阶段转向大规模应用阶段，在各行各业引发效率跃迁和商业模式变革，在能源、金融、制造等重点领域亦开始落地应 用，随之而来的安全问题也越来越被市场所重视。 智能体安全 应当从一 ...

Group 1: AI Development and Trends - The 2025 World Artificial Intelligence Conference (WAIC 2025) showcased a variety of discussions on the future of AI, emphasizing a shift from traditional conference formats to a "question-driven, deep dialogue" approach [1] - AI is breaking down traditional disciplinary barriers, particularly in fields like quantum physics, materials science, and biomedicine, leading to new research paradigms [3][4] - The integration of embodied intelligence and reinforcement learning is creating a new form of AI that closely resembles human intelligence, enabling real-world applications such as autonomous robots and self-driving cars [7][8] Group 2: AI in Life Sciences - AI is transforming life sciences by covering the entire research process, from pathology studies to molecular analysis, exemplified by systems like DeepMind's GNoME [5] - The development of digital twin brains is reshaping the understanding of the human brain, allowing for simulations of brain activity and predictions of neurological diseases [6] Group 3: AI Safety and Ethical Considerations - The rise of intelligent agents raises security concerns, with experts highlighting the need for a comprehensive protection system from design to deployment to ensure these agents are reliable partners [2] - Ethical considerations are paramount as technologies like digital twin brains challenge the boundaries of "thought privacy" and human consciousness [6][9]

Core Insights - The AI field is transitioning from the era of large models to the era of intelligent agents, with Ant Group's "Yitianjian" upgrading its security solutions to include AI agent safety assessment tools [1][2] - The upgraded features of "Yitianjian" include four core functions: agent alignment, MCP security scanning, intelligent agent security scanning, and zero-trust defense [1] - Over 70% of AI agent practitioners express concerns about issues such as AI hallucinations, erroneous decision-making, and data leaks, highlighting the safety challenges posed by intelligent agents [1] Company Insights - "Yitianjian" is a collaborative development between Ant Group and Tsinghua University, designed to ensure the safe and reliable operation of large model technologies [2] - The risk assessment agent of "Yitianjian" boasts an accuracy rate of over 96% and supports testing for intelligent agents across 11 industries [2] - The safety philosophy of the upgraded "Yitianjian" is based on the concept of "attack to promote defense," creating a comprehensive protection system for intelligent agents [2]