内容来源：https://www.wired.com/story/backchannel-anthropic-dispute-with-the-pentagon/

内容总结：

美国人工智能公司Anthropic近期与五美国防部的合作关系出现重大波折。该公司去年虽成为首家获美国政府批准涉足机密军事应用的主流AI企业，但本周事态急转直下——五角大楼正重新评估双方合作，一项价值2亿美元的合同可能被终止，理由是这家以安全为核心理念的AI公司拒绝参与某些致命性军事行动。

更严峻的是，美国国防部可能将Anthropic列为“供应链风险”企业，这意味着任何采用其AI技术的防务承包商都可能被排除在军方合作名单之外。五角大楼发言人肖恩·帕内尔对此回应称：“我们的合作伙伴必须愿意帮助作战人员在任何战斗中取胜，这关系到部队和美国人民的安全。”

这一事件折射出AI行业与国家安全需求间的深层矛盾。Anthropic曾公开主张加强AI监管，并明确禁止其Claude模型参与武器设计或自主杀伤系统，这种立场与当前美国政府推动AI军事化的政策产生直接冲突。国防部首席技术官埃米尔·迈克尔近期明确表示：“政府不会容忍AI公司限制军方在武器系统中使用AI的方式。”

值得注意的是，包括OpenAI、xAI和谷歌在内的多家AI企业正在积极争取更高级别的安全许可，以期承接军方机密项目。这种趋势引发行业内外担忧：当AI实验室竞相将其最先进技术投入军事应用时，是否会削弱其构建安全可靠AI系统的初衷？

科技史学家史蒂文·利维在评论中指出，当前AI军事化竞赛正在改变行业生态。仅数年前，国际社会还在认真探讨建立AI监管框架，如今这种讨论已逐渐被大国AI军备竞赛的现实所取代。更令人忧虑的是，当AI研发者将致命性功能嵌入系统时，可能从根本上改变这项技术的本质发展方向。

随着AI逐渐成为现代战争的核心要素，如何在技术创新、伦理约束与国家安全之间寻求平衡，已成为摆在所有AI开发者和政策制定者面前的严峻课题。正如阿西莫夫机器人第一定律所警示的：“机器人不得伤害人类”——在AI军事化浪潮中，这条基本原则正面临前所未有的挑战。

中文翻译：

去年，当Anthropic成为首家获美国政府批准用于涉密领域（包括军事应用）的大型AI公司时，这条新闻并未掀起太大波澜。但本周，第二波发展如炮弹般炸响：五角大楼正在重新评估与该公司的合作关系，其中包括一份价值2亿美元的合同。表面原因是这家注重安全的AI公司拒绝参与某些致命性军事行动。这个所谓的"战争部"甚至可能将Anthropic列为"供应链风险"——这枚通常预留给与中国等受联邦机构审查国家有商业往来的企业的"红字"标签，意味着五角大楼将不与任何在其国防工作中使用Anthropic AI的公司合作。五角大楼首席发言人肖恩·帕内尔向《连线》杂志证实Anthropic正处风口浪尖："我们的国家要求合作伙伴必须愿意帮助作战人员在任何战斗中取胜。归根结底，这关乎我们的军队和美国人民的安全。"这也是对其他公司的警示：目前持有国防部非涉密合同的OpenAI、xAI和谷歌，正在竭力争取各自的高级安全许可。

此事值得深入剖析。首先，Anthropic是否因抗议其AI模型Claude被用于推翻委内瑞拉总统马杜罗的突袭行动而遭受惩罚？（现有报道如此陈述，该公司予以否认）其次，Anthropic公开支持AI监管——这在行业内属于异类立场，且与政府现行政策相悖。但更令人不安的核心问题在于：政府对军事用途的需求是否会削弱AI本身的安全性？

研究者与高管们普遍认为AI是人类有史以来最强大的技术。几乎所有现存AI公司的创立前提，都是相信能够以防范大规模危害的方式实现通用人工智能或超级智能。xAI创始人埃隆·马斯克曾是约束AI的最主要倡导者——他联合创立OpenAI正是出于对这项技术可能被逐利企业掌控的深切担忧。

Anthropic则将自己定位为最注重安全的AI企业。该公司的使命是将防护机制深度植入模型，使恶意行为者无法释放AI最危险的潜能。艾萨克·阿西莫夫在机器人三定律中早已作出精辟阐述：机器人不得伤害人类，或因不作为而使人类受到伤害。即使当AI变得比任何人类都更聪明——这是AI领袖们笃信的未来——这些防护栏也必须屹立不倒。

因此，顶尖AI实验室争相将产品投入尖端军事和情报领域的现象显得尤为矛盾。作为首个获得涉密合同的主要实验室，Anthropic为政府提供了"专门为美国国家安全客户定制的Claude Gov模型系列"。该公司坚称此举未违背其安全标准，包括禁止使用Claude生产或设计武器。Anthropic首席执行官达里奥·阿莫迪曾明确表示不愿让Claude涉足自主武器或政府监控领域。但这可能与当前政府立场相冲突。国防部首席技术官埃米尔·迈克尔（前优步首席商务官）本周向记者表示，政府不会容忍AI公司限制军方在武器系统中使用AI的方式："当无人机群从军事基地涌出时，你有何应对方案？如果人类反应时间不够快……你该如何处置？"机器人第一定律在此显得如此苍白。

有观点认为有效的国家安全需要从最具创新力的公司获取顶尖技术。尽管几年前尚有科技公司对与五角大楼合作心存顾虑，但到2026年，它们大多已成为挥舞国旗的潜在军事承包商。我尚未听闻任何AI高管公开谈论其模型与致命武力的关联，但Palantir首席执行官亚历克斯·卡普却毫不讳言，甚至带着明显自豪宣称："我们的产品偶尔被用于终结生命。"

美国在与委内瑞拉这类国家交战时或许能肆意施展AI肌肉。但面对复杂对手时，各国将不得不积极部署自主的国家安全AI系统，从而引发全面军备竞赛。当涉及致命行为时，政府对于坚持划分"合法使用"界限的AI公司恐怕难有耐心（尤其是一个惯于重新定义法律来合理化战争罪行的政府）。五角大楼的声明已昭然若揭：若AI公司想与国防部合作，就必须承诺不惜代价争取胜利。

这种思维在五角大楼或许合理，却将构建安全AI的努力推向错误方向。若旨在创造不伤害人类的AI，同时开发致命武力版本无异于南辕北辙。就在数年前，各国政府与科技领袖还在认真探讨通过国际机构监督限制AI有害应用的方案。如今这类讨论已近乎绝迹。AI决定战争未来已成为既定事实。更令人恐惧的是，如果开发AI的企业与运用AI的国家未能谨慎约束这项技术，AI本身的未来可能更易与战争暴力同流合污。

我始终认为数字技术崛起是我们时代的主旋律。政客、政权甚至国家都可能更迭——但科技对人类社会的重塑不可逆转。当唐纳德·特朗普2016年首次当选总统时，我在题为《iPhone比特朗普更重要》的专栏中阐述了这一观点。2024年他再度当选时，我续写篇章论证AI是比总统更强大的混沌制造者。长远来看，科学终将超越特朗普。

如今这个理论似乎略显动摇。未来可能取决于谁掌控先进AI，以及他们如何塑造和利用它。当AI巨头们身披爱国主义外衣与五角大楼达成交易时，实情是他们正在向拒绝监管的政府与战争部门提供一种强大到令人恐惧且难以预测的技术。阿西莫夫若在世，会作何感想？

本文节选自史蒂文·利维的《反向通道》通讯专栏，过往文章可通过此处查阅。

英文来源：

When Anthropic last year became the first major AI company cleared by the US government for classified use—including military applications—the news didn’t make a major splash. But this week a second development hit like a cannonball: The Pentagon is reconsidering its relationship with the company, including a $200 million contract, ostensibly because the safety-conscious AI firm objects to participating in certain deadly operations. The so-called Department of War might even designate Anthropic as a “supply chain risk,” a scarlet letter usually reserved for companies that do business with countries scrutinized by federal agencies, like China, which means the Pentagon would not do business with firms using Anthropic’s AI in their defense work. In a statement to WIRED, chief Pentagon spokesperson Sean Parnell confirmed that Anthropic was in the hot seat. “Our nation requires that our partners be willing to help our warfighters win in any fight. Ultimately, this is about our troops and the safety of the American people,” he said. This is a message to other companies as well: OpenAI, xAI and Google, which currently have Department of Defense contracts for unclassified work, are jumping through the requisite hoops to get their own high clearances.
There’s plenty to unpack here. For one thing, there’s a question of whether Anthropic is being punished for complaining about the fact that its AI model Claude was used as part of the raid to remove Venezuela's president Nicolás Maduro (that’s what’s being reported; the company denies it). There’s also the fact that Anthropic publicly supports AI regulation—an outlier stance in the industry and one that runs counter to the administration’s policies. But there’s a bigger, more disturbing issue at play. Will government demands for military use make AI itself less safe?
Researchers and executives believe AI is the most powerful technology ever invented. Virtually all of the current AI companies were founded on the premise that it is possible to achieve AGI, or superintelligence, in a way that prevents widespread harm. Elon Musk, the founder of xAI, was once the biggest proponent of reining in AI—he cofounded OpenAI because he feared that the technology was too dangerous to be left in the hands of profit-seeking companies.
Anthropic has carved out a space as the most safety-conscious of all. The company’s mission is to have guardrails so deeply integrated into their models that bad actors cannot exploit AI’s darkest potential. Isaac Asimov said it first and best in his laws of robotics: A robot may not injure a human being or, through inaction, allow a human being to come to harm. Even when AI becomes smarter than any human on Earth—an eventuality that AI leaders fervently believe in—those guardrails must hold.
So it seems contradictory that leading AI labs are scrambling to get their products into cutting-edge military and intelligence operations. As the first major lab with a classified contract, Anthropic provides the government a “custom set of Claude Gov models built exclusively for U.S. national security customers.” Still, Anthropic said it did so without violating its own safety standards, including a prohibition on using Claude to produce or design weapons. Anthropic CEO Dario Amodei has specifically said he doesn’t want Claude involved in autonomous weapons or AI government surveillance. But that might not work with the current administration. Department of Defense CTO Emil Michael (formerly the chief business officer of Uber) told reporters this week that the government won’t tolerate an AI company limiting how the military uses AI in its weapons. “If there’s a drone swarm coming out of a military base, what are your options to take it down? If the human reaction time is not fast enough … how are you going to?” he asked rhetorically. So much for the first law of robotics.
There’s a good argument to be made that effective national security requires the best tech from the most innovative companies. While even a few years ago, some tech companies flinched at working with the Pentagon, in 2026 they are generally flag-waving would-be military contractors. I have yet to hear any AI executive speak about their models being associated with lethal force, but Palantir CEO Alex Karp isn’t shy about saying, with apparent pride, “Our product is used on occasion to kill people.”
The US might be able to flex its AI muscles with impunity when in combat with a country like Venezuela. But sophisticated opponents will have to aggressively implement their own versions of national security AI, with the result being a full-tilt arms race. The government will likely have little patience for AI companies that insist on carve-outs or lawyerly distinctions about what consists of “legal use” when a lethal practice is under question. (Especially a government that feels free to redefine the law to justify what many consider to be war crimes.) That Pentagon statement says it explicitly: If AI companies want to partner with the Department of Defense, they must commit to doing whatever it takes to win.
That mindset may make sense in the Pentagon, but it pushes the effort to create safe AI in the wrong direction. If you are creating a form of AI that won’t harm people, it’s counterproductive to also work on versions that deliver lethal force. Only a few years ago, both governments and tech executives were talking seriously about international bodies that might help monitor and limit the harmful uses of AI. You don’t hear that talk much any more. It’s a given now that the future of warfare is AI. Even more frightening, the future of AI itself might be more amenable to the kind of violence seen in warfare—if the companies that make it and the nations that wield it do not take care to contain the technology.
I have long believed that the major story of our times is the rise of digital technology. Politicians, regimes, and even countries may come and go—but tech’s remaking of humanity is irrevocable. When Donald Trump was first elected president in 2016, I spelled out this theory in a column called “The iPhone Is Bigger Than Donald Trump.” Upon his reelection in 2024, I wrote a sequel, arguing that AI was a bigger chaos agent than the president. In the long run, I argued, science trumps even Trump.
That theory now feels a little shakier. The future might hinge on who is in charge of advanced AI and how they shape and exploit it. While the lords of AI wrap themselves in patriotism and seek deals with the Pentagon, the fact is that they are supplying a fearsomely powerful and unpredictable technology to a government and a war department that rejects the idea of oversight. What would Asimov think?
This is an edition of Steven Levy’s Backchannel newsletter. Read previous newsletters here.