EVENT DOSSIER
As of August 22, 2026, several laboratories specializing in cutting-edge artificial intelligence technologies still do not disclose their specific containment strategies for preventing or curbing potentially失控 artificial intelligence models. According to TechCrunch, despite the growing attention in the industry to AI security, relevant institutions remain silent regarding how to implement risk control effectively, without revealing detailed technical approaches or implementation plans to the public.
Guidelight AI Standards 组织对 OpenAI、Anthropic、Meta、Google 和 xAI 五家领先实验室进行了安全评估，发现仅有少数机构发布了针对 rogue model（失控模型）的 containment plan（收容计划）。该计划旨在明确一旦 AI 试图颠覆人类控制时如何切断权限及完全关闭系统。OpenAI 得分最高，而 Anthropic 和 Meta 得分最低。此次评估基于各公司公开的计划，涵盖内部日志监控、异常行为熔断机制、第三方审计及具体失控应对方案等指标。随着具身智能（agentic AI）在企业系统中承担更多自主角色，以及加州和新 York 监管机构开始要求披露，这些发现揭示了实验室在运营风险与实际言论之间的差距。Guidelight 首席科学家 Ste…