AI News AI资讯 10h ago Updated 3h ago 更新于 3小时前 56

The AI off switch: How Anthropic’s export controls sparked a global AI sovereignty scramble AI 关机开关:Anthropic 的出口管制如何引发全球 AI 主权争夺战

Anthropic’s Fable 5 & Mythos 5 models were abruptly shut down globally on June 13, 2026. The US government cited a national security "jailbreak" vulnerability for the export control directive. Anthropic could not filter by nationality, so it disabled access for all users. The order came 4 days after the models’ general availability launch. European leaders decried the move as proof of dangerous dependence on US-controlled AI. 2026年6月13日,美国政府依据出口管制指令,要求Anthropic立即暂停其最强大AI模型Claude Fable 5和Mythos 5对所有外国用户的访问。 该禁令直接源于一个“越狱”漏洞争议,但深层背景是数月来Anthropic与美国政府在AI军事应用等问题上的冲突。 指令要求禁止“所有外国人”访问,迫使Anthropic在无法按国籍实时过滤的情况下,全面关停服务,包括其自身外籍员工。 欧洲多国反应强烈,视其为对“技术主权”的警钟,担忧关键数字基础设施可被单一国家政府“一键关闭”。 事件凸显了美国AI政策的内在矛盾:同一政府既指控Anthropic对本国构成供应链风险,又阻止其技术流向

85
Hot 热度
80
Quality 质量
75
Impact 影响力

Analysis 深度分析

TL;DR

  • Anthropic’s Fable 5 & Mythos 5 models were abruptly shut down globally on June 13, 2026.
  • The US government cited a national security "jailbreak" vulnerability for the export control directive.
  • Anthropic could not filter by nationality, so it disabled access for all users.
  • The order came 4 days after the models’ general availability launch.
  • European leaders decried the move as proof of dangerous dependence on US-controlled AI.

Key Data

Entity Key Info Data/Metrics
Claude Fable 5 Publicly released model class; safe for general use. Launch Date: June 9, 2026
Claude Mythos 5 More powerful sibling; access restricted to "Project Glasswing" partners. Restricted Access since April 2026
Anthropic Received export control directive at 5:21 PM ET on June 12, 2026. Compliance required "abruptly disabling" access for all customers.
US Government Directive issued by Commerce Secretary Howard Lutnick to CEO Dario Amodei. Cites specific "jailbreak" vulnerability for Fable 5.
David Sacks Co-chair of President’s Council of Advisers on Science and Technology. Stated Amodei refused to fix vulnerability or pull model.
Amazon Major Anthropic investor; provided security concerns to US Treasury. Researchers found prompts could aid cyberattacks.
European Commission Confirmed examining the fallout from the shutdown. Published Technological Sovereignty Package on June 3, 2026.

Deep Analysis

The sudden, unilateral shutdown of Anthropic’s most powerful AI models by the U.S. government isn’t a policy dispute—it’s a geopolitical seizure in real-time. The official narrative about a critical "jailbreak" vulnerability feels like the thinnest of pretexts. Anthropic’s rebuttal, that the technique merely allows limited code review, a capability standard across the industry including with OpenAI’s GPT-5.5, exposes the move as politically, not technically, motivated. This isn’t about a flaw; it’s about control.

The core contradiction is staggering and deliberately maintained. Within the same year, the Trump administration has simultaneously labeled Anthropic a "Supply-Chain Risk to National Security" too dangerous for federal use and then weaponized export controls to deem its technology too dangerous for global use. This isn’t hypocrisy; it’s a strategy to discipline a corporation. Anthropic drew a red line against mass surveillance and autonomous weapons, and the Pentagon’s response was to try and cripple it commercially. The export order is the knockout punch, reframing a domestic blacklisting as a global security imperative.

The incident lays bare the fiction of "corporate sovereignty" in the age of frontier AI. A U.S. government email, sent in an afternoon, turned off a tool embedded in companies, research institutions, and public services across multiple continents. The fact that Anthropic, a company with safety as its brand ethos, was forced to sever access for its own foreign-born employees underscores the brutal, extraterritorial reach of this authority. It’s a stark reminder that in the current paradigm, there is no "global AI"—there are American AI systems, temporarily licensed to the world at the pleasure of Washington.

The European reaction is therefore not about cybersecurity; it’s a visceral panic about sovereignty. The timing is exquisitely bad for the EU, coming nine days after its own Technological Sovereignty Package. The shutdown validated the most paranoid fears of European digital strategists overnight. The commentary from Finland and France is correct: building a digital future on infrastructure that can be toggled off by a foreign capital is not a strategy, it’s a vulnerability. This event will single-handedly pour rocket fuel on initiatives for European model development and sovereign cloud infrastructure. It moves the debate from abstract principle to urgent survival.

Amazon’s alleged role is the grim cherry on top. Its reportedly providing intelligence about Fable 5’s cyber risks to the Treasury showcases the new corporate-state intelligence nexus. Tech giants are not just partners to the state; they are informants, leveraging their unique access to shape policy against competitors or uncooperative firms. This creates a chilling feedback loop where the most capable models, closely scrutinized by their own ecosystems, become the most vulnerable to regulatory attack.

Ultimately, Anthropic is trapped in a classic security dilemma. It built something deemed too powerful, then tried to impose ethical guardrails that conflicted with the national security state’s objectives. The response wasn’t to debate the ethics, but to assert absolute jurisdictional control. The message to every AI lab globally is clear: build what we allow, and never constrain how we use it. The frontier of AI development has just been forcibly relocated to a more compliant, and more dangerous, legal landscape.

Industry Insights

  1. Sovereign AI initiatives will accelerate dramatically in Europe, the Middle East, and Asia as governments seek to eliminate single-point-of-failure dependencies.
  2. AI labs will preemptively develop geopolitical risk mitigation strategies, including geographic access segmentation and diversified model licensing structures.
  3. Expect increased regulatory scrutiny on the dual-use capabilities of all frontier models, with governments demanding "kill switch" access as a condition of deployment.

FAQ

Q: Why did the US government shut down the models?
A: The government cited a specific "jailbreak" vulnerability in Fable 5 that it deemed a national security threat, though Anthropic disputed its severity and uniqueness.

Q: How could the US government disable models for people outside the US?
A: The order applied to any foreign national anywhere. Since Anthropic couldn't filter users by nationality in real-time, it had to disable access for all customers to comply.

Q: What does this mean for the future of global AI development?
A: It confirms that powerful AI models are now considered geopolitical assets and weapons. Development and access will increasingly be dictated by national security alliances, not just commercial or scientific merit.

TL;DR

  • 2026年6月13日,美国政府依据出口管制指令,要求Anthropic立即暂停其最强大AI模型Claude Fable 5和Mythos 5对所有外国用户的访问。
  • 该禁令直接源于一个“越狱”漏洞争议,但深层背景是数月来Anthropic与美国政府在AI军事应用等问题上的冲突。
  • 指令要求禁止“所有外国人”访问,迫使Anthropic在无法按国籍实时过滤的情况下,全面关停服务,包括其自身外籍员工。
  • 欧洲多国反应强烈,视其为对“技术主权”的警钟,担忧关键数字基础设施可被单一国家政府“一键关闭”。
  • 事件凸显了美国AI政策的内在矛盾:同一政府既指控Anthropic对本国构成供应链风险,又阻止其技术流向盟友。

核心数据

实体 关键信息 数据/指标
政府指令生效日期 出口管制指令暂停Fable 5和Mythos 5访问 2026年6月13日
模型公开发布日期 Fable 5和Mythos 5向公众发布 2026年6月9日
从发布到下架 模型从公开可用到被全面关停的时间间隔 4天
收到指令时间 Anthropic称收到政府指令的具体时间 美国东部时间6月12日下午5:21
签发人 商务部长 Howard Lutnick
相关企业风险 亚马逊CEO曾向政府官员提及Fable 5的安全风险 Andy Jassy
反对企业估值 Anthropic在诉讼中警告的潜在收入损失 数亿美元
监管行动日期 欧盟发布《技术主权一揽子计划》的日期 2026年6月3日

深度解读

这起事件的本质,不是一次技术安全响应,而是一场披着“漏洞”外衣的、蓄谋已久的政治围剿。所谓“越狱”漏洞的争议,更像是美国政府为早已准备好的弹药找到的一个响亮扳机。

让我们拆解这个“矛盾”的叙事。政府说Anthropic的模型太危险,不能给外国人用,甚至其内部的外国人都不能碰;但就在不久前,政府还说这家公司是“供应链安全风险”,自己都不要用。这像不像一个人一边大喊“你有毒!离我远点!”,一边又对邻居嘶吼“谁也不准碰我的毒药!”。Anthropic恰恰卡在了这中间:它被指控同时“对内有毒”和“对外危险”,成了AI版的“里外不是人”。

这暴露了美国AI政策核心的、无法调和的分裂。一部分政府力量(如商务部、国家安全顾问)将先进AI视为必须严格控制的战略武器;另一部分力量(以国防部为代表)则可能因为Anthropic不愿其技术被用于大规模监控或全自动武器而将其“打入冷宫”。Anthropic的“AI安全”品牌,在此刻与其说是一张护身符,不如说成了一个政治靶子——它既因坚持某些安全原则而触怒了军方,又因其模型的强大能力而成为了地缘博弈的焦点。

更值得玩味的是亚马逊的角色。作为投资方,其CEO向政府“吹风”安全风险,这一举动充满了算计。是真忧心安全,还是试图在政治风暴中为自己与Anthropic的关系寻找一个切割点或护身符?这为科技巨头与政府之间那种既亲密又猜忌的共生关系,增添了又一个经典注脚。

而整个事件最令人不寒而栗的启示,不是“越狱”漏洞本身,而是展示了一个可怖的先例:一个国家政府,可以通过一纸行政命令,在一个下午关停一个已嵌入全球企业与研究机构运营的、关键的数字生产工具。 欧洲和加拿大的恐慌是真实的。芬兰政客的反应——“不能继续依赖一个可以被外国政府一夜之间关闭的准入渠道”——道破了天机。这无关漏洞,关乎主权。美国用行动证明,即使是其盟友,在AI这个新战场上,也可能随时被“拔掉网线”。这记耳光,打得欧洲“技术主权”的口号瞬间无比具体而沉重。

Anthropic正处在一个荒诞的十字路口:它可能既无法服务美国政府,也无法服务全球市场,同时还要应对投资者和公众对其“安全公司”承诺的质疑。这不是一家公司的危机,这是全球AI产业在“美国制造”技术霸权与全球协作需求之间,一道越来越深的裂谷的首次公开塌方。

行业启示

  1. “技术主权”从政治口号变为企业的紧急采购清单:全球企业将加速评估并寻求非美国或开源的AI模型作为备选,以降低单一国家政策突变带来的运营风险。
  2. AI安全的定义权已成为地缘政治武器:所谓“安全风险”的认定标准可能不再基于客观技术分析,而服务于特定的外交与战略目标,这将使AI合规变得极度复杂和不确定。
  3. 开源与去中心化AI架构的吸引力将激增:为规避集中式控制风险,对分布式、可本地化部署的AI解决方案的需求,将从技术偏好上升为战略必需。

FAQ

Q: 这意味着我不能再使用Claude AI了吗?
A: 不一定。此禁令针对的是Anthropic最强大的Fable 5和Mythos 5模型,且基于国籍进行限制。但Anthropic的其他早期版本模型或未来针对美国公民开放的版本可能不受影响,具体取决于公司后续的合规措施与模型发布策略。

Q: 其他美国AI公司(如OpenAI)会面临同样情况吗?
A: 理论上有可能。此事开创了一个危险先例:只要政府援引“国家安全”,即可对任何具备战略重要性的AI模型实施跨境使用限制。其他公司的尖端模型未来也可能因类似原因被施加出口管制。

Q: 对开源AI社区意味着什么?
A: 这极大地强化了开源AI的战略价值。当闭源模型可被政府“一键关闭”时,开源代码、开放权重和开放基础设施为全球提供了至关重要的“逃生通道”,但同时也可能面临更严格的出口管制审查。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

政策 政策 监管 监管 安全 安全
Share: 分享到: