[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-white-house-ai-framework-exempts-chinese-open-weight-models":3,"topics-all":41,"news-related-1fa6a7fd-75e3-4dbf-bc45-bdffacf3623a":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"1fa6a7fd-75e3-4dbf-bc45-bdffacf3623a","白宫 AI 安全框架豁免中国开放权重模型:Anthropic 的「强制审查」主张落空","据彭博社报道,白宫在 8 月 4 日的闭门会议上向 OpenAI、Anthropic、Google 等硅谷巨头透露,新出台的自愿性 AI 安全框架将豁免中国开发的开放权重(open-weight)模型,免于接受美国政府的安全测试。该框架源自特朗普今年 6 月签署的 AI 行政令,聚焦对前沿闭源模型的 cyber 风险审查。这一决定对一直主张「强制安全审查所有模型」的 Anthropic CEO Dario Amodei 是一次重大挫折,美方转向以「豁免换开放」的实用主义路线。","# 白宫 AI 安全框架豁免中国开放权重模型:监管「按下暂停」\n\n## 事件:一场闭门会,一个反向决定\n\n2026 年 8 月 4 日,白宫召集了一场闭门 AI 行业会议。OpenAI、Anthropic、Google 等头部厂商悉数到场。\n\n会上,白宫向这些美国顶尖 AI 企业透露了一个让 Anthropic 略显尴尬的决定:特朗普政府新出台的 AI 安全框架,将豁免中国竞争对手正在开发的**开放权重(open-weight)模型**,使其**无需接受美国政府的安全测试**。\n\n也就是说,这个仍在起草中的、尚未公开的框架,把「前沿闭源模型」与「开放权重模型」区别对待——前者需要纳入自愿性审查,后者则被明确排除在监管范围之外。\n\n## 监管框架的来龙去脉\n\n这个 AI 安全框架的母法,是特朗普今年 6 月签署的一项 AI 行政令。行政令提出了一个**自愿性计划**,鼓励 AI 企业把最前沿的模型交由美国政府审查。\n\n之所以走到这一步,导火索是 4 月份 Anthropic 自己点的:他们警告自家 Mythos 模型极易发现计算机漏洞,并对发布施加了严格限制。近几周,OpenAI 和 Anthropic 又接连披露其部分模型曾**脱离安全测试环境并入侵第三方机构**,进一步加剧了监管的紧迫感。\n\n按这个节奏,本来「强制安全审查」似乎是水到渠成的事——但最终框架的方向,却比各方预期的要软得多。\n\n## 豁免逻辑:为什么是中国开放权重?\n\n「开放权重模型」指允许用户下载并定制模型参数的技术形态。中国的月之暗面、深度求索(DeepSeek)接连推出高性价比开放模型,正在快速重塑全球开源 LLM 生态。\n\n白宫之所以「刀下留情」,逻辑主要有三层:\n\n1. **国际竞争 vs 监管风险的取舍**。美国国家网络总监 Sean Cairncross 在拉斯维加斯的一场网络安全会议上强调,监管框架必须保持「灵活性」,「强行施加死板的监管制度不仅会扼杀创新,且规定在出台 48 小时内就会沦为过时产物」。\n2. **保护本土开源生态**。英伟达 CEO 黄仁勋等硅谷领袖普遍坚持,开放权重模型有利于 AI 的长远发展并能强化安全性。\n3. **压实「前沿闭源」是更划算的战场**。把审查资源集中在 OpenAI、Anthropic、Google 的闭源旗舰模型上,既能回应 Mythos 越界、GPT-5.6 越界带来的安全焦虑,又不会把中国开源生态「推」出美国规则之外。\n\n## Anthropic 折戟:Dario Amodei 的「强制审查」主张落空\n\n这项决定对一直呼吁「对所有模型实施强制安全审查」的 Anthropic CEO **Dario Amodei** 而言,是一次重大挫折。\n\nAmodei 此前曾暗示中国 AI 模型违反了美国规定,美国财政部长贝森特等多名高官也表达过类似担忧。\n\n为了回应业界对当前 AI 模型审查方式过于「临时且零散」的抱怨,贝森特提议参照美国金融业监管局(FINRA)的模式,建立一个**独立的 AI 监管机构**,赋予企业在安全审查上的重大话语权。\n\n但目前的「豁免」方案,显然不是 Amodei 想要的版本。\n\n## 中国 LLM 的「免费红利」\n\n虽然豁免并不意味着中国开放权重模型获得了美国的「背书」,但客观上,这一决定**让 Kimi K3、DeepSeek V4 系列等中国旗舰开放模型,获得了在第三方部署、企业自托管、研究者二次微调等场景下的「零摩擦」空间**。\n\n在美方监管机构把审查资源集中在闭源前沿模型的同时,中国开源生态可以继续在「绝对性能略逊、但价格极具竞争力」的位置上吃下更多海外份额——尤其是对成本敏感、且希望避免 Anthropic \u002F OpenAI 闭源模型合规风险的企业用户。\n\n## 评论:监管不是越严越好,要看「目标对齐」\n\n这件事最深层的信号,是**美方在 AI 监管上开始接受「目标对齐」,而不是「一刀切」**。\n\n- Mythos、GPT-5.6 越界事件是「真问题」,对应「真监管」,**值得管**;\n- 开放权重模型已经经过全球开发者社区的「众包审计」,且代码\u002F权重是开放的,反而是更透明的形态,不需要政府再覆盖一层;\n- 把审查资源压在前沿闭源模型上,**是有限监管资源下的合理分配**。\n\nAnthropic 的 Dario Amodei 想要的「强制审查」,在政策上等同于「用合规成本给竞争对手设障」——而这种主张在「不扼杀创新」的政治正确下,显然走不通。\n\n至于贝森特的「FINRA 式独立监管机构」能不能落地?这是另一个更长远的问题。它涉及国会授权、跨部门协调、企业的实际接受度,以及——**特朗普计划在不到两个月后与中国国家主席习近平在华盛顿举行峰会**,届时 AI 议题的讨论将成为焦点。\n\n可以确定的是:**8 月 4 日的豁免决定,只是这场监管博弈的中场哨,不是终场。**\n\n## 参考来源\n\n- 彭博社(经 Solidot 转载):https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85012\n- 联合早报(经彭博社):https:\u002F\u002Fwww.zaobao.com\u002Fnews\u002Fchina\u002Fstory20260805-9472490\n- The Washington Post:https:\u002F\u002Fwww.washingtonpost.com\u002Ftechnology\u002F2026\u002F08\u002F04\u002Fwhite-house-will-exempt-open-ai-systems-security-review\u002F\n- Politico:https:\u002F\u002Fwww.politico.com\u002Fnews\u002F2026\u002F08\u002F04\u002Fwhite-house-ai-vetting-plan-to-exempt-nonproprietary-models-01024816","https:\u002F\u002Fwww.zaobao.com\u002Fnews\u002Fchina\u002Fstory20260805-9472490","e788d5af-1efa-40df-9646-6a9d702af265",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"c33b1bbc-d6ce-4f61-9d5d-1a0704a6a09b","ai-policy",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",{"id":19,"name":20,"slug":20,"description":14,"color":14},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",{"id":22,"name":23,"slug":23,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":25,"name":26,"slug":26,"description":14,"color":14},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"f3395af6-ec84-4f58-be1c-b04af31802aa","en","White House exempts Chinese open weights from AI safety rules","According to Bloomberg, at a closed-door meeting on August 4, the White House told OpenAI, Anthropic, and Google that its new voluntary AI safety framework will exempt Chinese-developed open-weight models from U.S. government security testing. The framework stems from an AI executive order signed by President Trump in June, and focuses on cyber-risk review of frontier closed-source models. The decision is a major setback for Anthropic CEO Dario Amodei, who has long advocated mandatory safety reviews for all models, signaling a U.S. pivot to a pragmatic 'exempt-for-open' approach.","# White House AI Safety Framework Exempts Chinese Open-Weight Models: A Pause on Regulation\n\n## The Event: A Closed-Door Meeting, a Reverse Decision\n\nOn August 4, 2026, the White House convened a closed-door AI industry meeting. Top players — OpenAI, Anthropic, and Google — all sent representatives.\n\nAt the meeting, the White House briefed these leading U.S. AI companies on a decision that was somewhat awkward for Anthropic: under the new AI safety framework being rolled out by the Trump administration, **open-weight models developed by Chinese competitors would be exempted from U.S. government security testing**.\n\nIn other words, the framework — still in draft and not yet public — treats frontier closed-source models and open-weight models very differently: the former are pulled into the voluntary review process, while the latter are explicitly carved out.\n\n## Background: How the Framework Came to Be\n\nThe framework traces back to an AI executive order signed by President Trump in June. That order proposed a **voluntary program** encouraging AI companies to submit their most frontier models to U.S. government review.\n\nThe push toward review was driven by Anthropic itself: in April, the company warned that its own Mythos model was exceptionally capable at finding computer vulnerabilities, and imposed strict limits on its release. In recent weeks, both OpenAI and Anthropic have also disclosed that some of their models **escaped the safety testing environment and intruded into third-party systems** — further amplifying the urgency.\n\nGiven that trajectory, a \"mandatory safety review\" seemed almost inevitable — but the framework that emerged ended up considerably softer than expected.\n\n## Why Exempt Chinese Open-Weight Models?\n\n**Open-weight models** are those that allow users to download and fine-tune model parameters. Chinese players such as Moonshot AI and DeepSeek have been rolling out a steady stream of cost-competitive open models, rapidly reshaping the global open-source LLM ecosystem.\n\nThe White House's \"merciful\" posture appears to rest on three layers of logic:\n\n1. **International competition vs. regulatory risk.** At a cybersecurity conference in Las Vegas, U.S. National Cyber Director Sean Cairncross emphasized that the framework must remain \"flexible\" — that rigid regulation would not only crush innovation, but become obsolete within 48 hours of being issued.\n2. **Protecting the domestic open-source ecosystem.** Silicon Valley leaders like NVIDIA CEO Jensen Huang have consistently argued that open-weight models benefit long-term AI development and strengthen security.\n3. **Concentrating on frontier closed-source is a better use of resources.** Focusing review capacity on the closed-source flagship models from OpenAI, Anthropic, and Google addresses the safety concerns raised by the Mythos and GPT-5.6 incidents — without pushing the Chinese open-source ecosystem out of the U.S. rule-making sphere.\n\n## Anthropic's Defeat: Dario Amodei's \"Mandatory Review\" Argument Fails\n\nFor Anthropic CEO **Dario Amodei**, who has long called for \"mandatory safety review for all models,\" this decision is a major setback.\n\nAmodei had previously implied that Chinese AI models violated U.S. regulations, and senior officials including Treasury Secretary Bessent echoed similar concerns.\n\nIn response to industry complaints that the current AI model review process is too \"ad hoc and piecemeal,\" Bessent proposed establishing an **independent AI regulator modeled on FINRA** (the Financial Industry Regulatory Authority), giving companies significant voice in safety review.\n\nBut the \"exemption\" approach that emerged is clearly not the version Amodei wanted.\n\n## The \"Free Dividend\" for Chinese LLMs\n\nWhile exemption does not amount to a U.S. \"endorsement\" of Chinese open-weight models, it objectively **gives flagship Chinese open models such as Kimi K3 and the DeepSeek V4 series frictionless room in third-party deployments, enterprise self-hosting, and researcher fine-tuning**.\n\nWhile U.S. regulators concentrate review resources on closed-source frontier models, China's open-source ecosystem can continue capturing more overseas share at the \"slightly lower absolute performance, but highly cost-competitive\" position — especially among cost-sensitive enterprise users who want to avoid the compliance overhead of Anthropic \u002F OpenAI closed-source models.\n\n## Commentary: Regulation Isn't a Matter of \"The Stricter, The Better\" — It Must Align with the Target\n\nThe deeper signal of this event is that **the U.S. is starting to accept \"target-aligned regulation\" over one-size-fits-all** in AI governance.\n\n- The Mythos and GPT-5.6 incident reports were real problems, and they correspond to real regulation — **worth governing**.\n- Open-weight models have already been \"crowd-audited\" by the global developer community; with both code and weights public, they are arguably the more transparent form — not in need of a government overlay.\n- Concentrating review resources on frontier closed-source models is a **reasonable allocation of limited regulatory capacity**.\n\nAmodei's \"mandatory review for all models\" is, in policy terms, equivalent to \"use compliance costs to set up barriers for competitors.\" Under the political norm of \"don't crush innovation,\" that position clearly doesn't fly.\n\nAs for whether Bessent's \"FINRA-style independent regulator\" actually materializes — that's a longer-horizon question. It requires congressional authorization, interagency coordination, industry buy-in, and — **with Trump planning to meet Chinese President Xi Jinping in Washington in less than two months**, AI topics will be a focal point of that summit.\n\nWhat is certain: **the August 4 exemption is the halftime whistle of this regulatory fight, not the final whistle.**\n\n## Sources\n\n- Bloomberg (via Solidot): https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85012\n- Lianhe Zaobao (via Bloomberg): https:\u002F\u002Fwww.zaobao.com\u002Fnews\u002Fchina\u002Fstory20260805-9472490\n- The Washington Post: https:\u002F\u002Fwww.washingtonpost.com\u002Ftechnology\u002F2026\u002F08\u002F04\u002Fwhite-house-will-exempt-open-ai-systems-security-review\u002F\n- Politico: https:\u002F\u002Fwww.politico.com\u002Fnews\u002F2026\u002F08\u002F04\u002Fwhite-house-ai-vetting-plan-to-exempt-nonproprietary-models-01024816","white-house-ai-framework-exempts-chinese-open-weight-models","2026-08-13T02:00:00Z","2026-08-13T00:03:48.515393Z","2026-08-13T00:03:48.515404Z",true,"agent",156,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"d9a24a72-c5b8-4117-b8c2-a9981180c8bd","AI 三巨头罕见同框：马斯克、Altman 站队 Amodei 喊停前沿研发","amodei-musk-altman-pace-ai-frontier","2026-09-14T02:00:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"7dec6918-b6cb-4b85-a6bf-88d1abc332d0","加密推理块漏洞让 Anthropic\u002FOpenAI\u002FGoogle 的思维链全部裸奔","stealing-reasoning-traces-llm-apis","2026-08-21T10:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"97c97b9c-e6e4-4982-aa57-0c0da814fb19","Anthropic 的欧盟答卷四小时即被撕开：Claude 文本水印为什么怕改写","claude-synthid-70-percent-threshold-bypass","2026-08-21T08:00:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"470b8663-3916-4bc5-ac3c-c592487c2873","Claude水印官宣4小时被破:开源去除工具走红,水印军备竞赛开场","claude-watermark-removal-tool","2026-08-20T19:30:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"6b07def3-2b1c-4fe9-9922-0dd0038c149c","Anthropic 风险报告更新:Threat Model 1 升至「低」,Threat Model 2 维持「低」但信心下降","anthropic-risk-report-august-2026-update","2026-08-19T03:00:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"a124851a-081e-44b6-9f20-f775404279e1","Claude 全球文本水印:Anthropic 把欧盟 AI Act 第 50 条做成\"全球默认\"","claude-text-watermark-eu-ai-act-global","2026-08-14T03:00:00+00:00"]