全球AI:2026年5月25日重要事件回顾
今天的《每日AI世界简报》汇集了来自世界关键地区的关于人工智能的最新消息。重点关注商业应用、监管、安全以及AI模型的发展。 欧洲 教宗利奥呼吁制定全球规则“解除”人工智能的威胁 - The Irish Times 教宗利奥呼吁制定全球规则“解除”人工智能的威胁 The Irish Times 重要性: 有必要关注该信息对市场、监管和AI用户的影响。 来源: Google News AI Europe (25.05.2026) WHO称AI在欧盟医疗保健领域的应用日益普及 - DD News WHO称AI在欧盟医疗保健领域的应用日益普及 DD News 重要性: 有必要关注该信息对市场、监管和AI用户的影响。 来源: Google News AI Europe (25.05.2026) 教宗利奥十四呼吁在宏伟宣言中对AI进行“强力”监管 - France 24 教宗利奥十四呼吁在宏伟宣言中对AI进行“强力”监管 France 24 为什么重要: 有必要关注该信息对市场、监管和人工智能用户的影响。 来源: Google News AI Europe (25.05.2026) 医疗领域的 AI:塑造未来 - CEPS 医疗领域的 AI:塑造未来 CEPS 为什么重要: 有必要关注该信息对市场、监管和人工智能用户的影响。 来源: Google News AI Europe (24.05.2026) 北美 归因错误鸿沟:当记忆投毒看起来像智能体 AI 系统中的模型故障时 arXiv:2605.22842v1 Announce Type: cross Abstract: Multi-agent AI pipelines typically assume that agent misconduct originates from model misalignment.
We identify a structural failure in this assumption, the \emph{Misattribution Gap}, where memory-layer attacks produce behaviors indistinguishable from model failure, causing defenders to apply the wrong remediation.
We formalize \emph{Semantic Norm Drift} (SND) as a third path to agent misconduct, distinct from emergent misalignment and collusion.
In SND, a policy-formatted document enters a shared vector store through normal uploads and later reappears as trusted system context after provenance is lost through a Trust Laundering Chain.
Across 64 documented failures, attribution systems consistently blamed the model.
Four safety classifiers, including one trained on memory poisoning, produced zero detections across 510 checkpoints.
In 59 o 重要性: 值得观察此信息对市场、监管和AI用户的影响。 来源: arXiv AI (25.05.2026) SafeHarbor:用于LLM智能体安全的层次化记忆增强护栏 arXiv:2605.05704v2 Announce Type: replace-cross Abstract: 基础模型领域的最新进展已将大型语言模型(LLMs)从被动的对话系统转变为具备推理和工具执行能力的自主智能体。虽然这些能力释放了巨大的实际价值,但也引入了新的安全风险,因为攻击者可能会操纵智能体在现实环境中执行有害行为。现有的防御策略可以减轻此类威胁,但往往难以平衡安全性与实用性,导致对良性用户请求过度拒绝。为了缓解这种权衡取舍,我们提出了 SafeHarbor,这是一个旨在为 LLM 智能体建立精确决策边界的新颖框架。与静态指南不同,SafeHarbor 通过增强的对抗生成提取上下文感知的防御规则。我们设计了一个本地分层记忆系统用于动态规则注入,提供 重要性: 有必要关注该信息对市场、监管和人工智能用户的潜在影响。 来源: arXiv AI (25.05.2026) 人机循环多智能体呼吸机决策支持与上下文老虎机偏好学习 arXiv:2605.23320v1 Announce Type: new Abstract: 呼吸机决策支持系统需要进行顺序决策,以跟踪不断变化的生理和疾病轨迹,同时尊重安全边界和临床医生特定的调整风格。基于规则的方法很少能实现个性化泛化,而端到端的强化学习或单一大型语言模型系统仍然难以控制和审计。我们提出了呼吸机决策支持系统(VDSS),这是一个人在回路中的多智能体框架,通过合同驱动的结构化接口协调模块化的决策组件,并产生可追溯的证据供审查。VDSS使用上下文老虎机执行在线偏好适应,在每个调整周期从最终接受的决策中更新临床医生特定的偏好,并利用这些偏好来指导后续建议。结构化拒绝反馈触发 为什么重要: 有必要观察该信息对市场、监管和人工智能用户的影响。 来源: arXiv AI (25.05.2026) MemAudit:通过因果归因和结构异常检测对毒化代理记忆进行事后审计 arXiv:2605.23723v1 Announce Type: new Abstract: Large language model agents increasingly rely on persistent memory to store past interactions, retrieve relevant demonstrations, and improve long-horizon task execution.
However, this memory mechanism also creates a practical security vulnerability: an adversarial user may inject malicious records into the agent's memory through ordinary interaction, and these records can later be retrieved to steer the agent's reasoning and actions.