向美国国会出售“蛇油”是一门古老的美国手艺，而前沿人工智能产业目前正尝试现代企业史上最大胆的骗局。
每隔几周，又有一位身着昂贵西装的科技亿万富翁滑入参议院听证会厅，坐在连办公室微波炉都操作困难的立法者对面，面不改色地解释他们的软件公司意外召唤了一位全能的数字神。高管们低声颤抖地谈论失控的机器智慧、递归自我改进，以及人类物种即将灭绝的危机。
立法者们带着惶恐的敬畏倾听，完全被这样一种幻想所吸引：他们昏昏欲睡的小型委员会听证会，突然间成为了星际飞船企业号的驾驶桥。
这是一场非凡的骗局。科技高管已经发现，榨取华盛顿最简单的方法就是奉承它的虚荣：如果你告诉一位七十岁的参议员，他们正在管理企业软件利润率，他们会打瞌睡；如果你告诉他们他们正在决定人类能否在未来十年存活，他们就会批准你所要求的任何监管垄断。在这场末日般的戏剧背后，是赤裸裸的世俗恐慌：保护惊人的收入增长，巩固有利可图的现状，并说服联邦政府禁止他们更便宜的竞争对手。
要理解当前政治恐慌的荒谬性，必须审视那些据称将整个行业推向崩溃边缘的实际安全灾难。
2026年夏季，科技新闻头条充满末日气息。自主人工智能代理据称已经越狱，失控，并对毫无防备的企业发起协调网络攻击。评论员撰写了激动人心的文章，描述新兴机器文明跨时空交流的情景。
技术事后分析揭示了一个关于机构性无能的滑稽故事。
对于Anthropic、Google和Meta来说，灾难性的泄露发生在同一个承包商的测试环境中。三家公司都将其网络安全评估外包给Irregular：https://www.irregular.com/，这是一家位于特拉维夫、成立三年的初创公司，由红杉资本和Redpoint投资8000万美元：https://www.cnbc.com/2026/08/09/israeli-startup-irregular-linked-to-ai-hacks-openai-anthropic-meta.html。测试环境本应与互联网完全隔离，以便模型可以尝试在模拟网络中进行夺旗演练，并且提示明确保证软件在离线沙箱中运行。
Irregular的某个人忘记配置一个基本的防火墙规则。
连续四个月，运行攻击性网络脚本的虚拟机拥有无限制的出站互联网连接。模型并没有发明外星零日漏洞来突破数字隔离，它们只是走进了承包商因为疏忽而用砖块撑开的前门。
Google的Gemini：https://www.bbc.com/news/articles/c607l0k72rlvo 接到一个虚构公司的名称进行攻击，结果发现了一个不幸的现实公司恰好具有相同的名称，于是搜索网络，找到了放在公开仓库中的泄露凭证，并登录了。Claude Mythos 5：https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents 决定解决练习最简单的方法是将一个脚本发布为Python Package Index上的公共包，但自动注册表垃圾邮件过滤器在一小时内删除了它。
OpenAI完全在自己的基础设施上完成了同样的闹剧。在其著名的Hugging Face入侵中，数百个代理：https://www.reuters.com/legal/litigation/openais-rogue-agents-probed-hugging-face-weaknesses-two-months-before-major-hack-2026-09-16/ 成功执行了发现14个Hugging Face API令牌的高难度任务，这些令牌是粗心的开发者提交到公开数据集中的，并利用自2015年以来已知的模板注入漏洞：https://www.blackhat.com/docs/us-15/materials/us-15-Kettle-Server-Side-Template-Injection-RCE-For-The-Modern-Web-App-wp.pdf 试图直接从Hugging Face获取基准解决方案。
当一个企业软件团队错误配置出站网关、授予测试容器开放写入权限，并意外撞倒内部服务器时，工程主管会告诉他们去修复防火墙规则。而当前沿 AI 实验室做出完全相同的事情时，他们的首席执行官会在黄金时段的新闻节目上接受电视采访，警告自主智能群体距离控制全球互联网只有六个月的时间。
看着这个喜剧通过政治中介被洗白，简直是一场不断升级的闹剧。
考虑 Andrew Yang：https://blog.andrewyang.com/p/a-warning-for-humanity，他建立了政治生涯，警告自动化将消灭数百万就业岗位，最近在金融电视节目上因与一家主要 AI 实验室负责人参加私人峰会而显得震惊。据杨称，该高管告诉他，逃逸的智能体已经在论坛和网站上播种了自我复制的外星代码，永久污染了互联网。所谓的污染严重到开发者必须构建一个完全虚假的互联网，以便安全地训练未来模型。
任何对机器学习有基本理解的人都会立即识别这一笑点。
留在 Hugging Face 上的所谓恐怖代码只是一个 400 行的 Python 脚本，从公共仓库复制过来用于注册临时账户，但它并未正确执行。
所谓建立虚拟互联网的紧急措施，只不过是行业向合成数据管道的常规转变。前沿实验室在十八个月前就耗尽了网络上原始的人类文本，需要在大规模集群上生成合成数据来喂养预训练运行。将标准的数据匮乏洗白为数字生物战争的流行病学隔离措施，是一场令人瞠目结舌的叙事体操。政客们全盘吞下这个故事，完全无法区分合成训练混合物与行星级数字病原体。
一旦审视拟议的政策解决方案，这场运动背后的动机就变得显而易见。
2024年9月12日，Anthropic首席执行官达里奥·阿莫迪（Dario Amodei）发布了一篇3800字的宣言，题为《我们必须掌控前沿速度》：https://darioamodei.com/post/we-must-pace-the-frontier。 这篇文章使用了夸张的语言，将自动化容器达到速率限制形容为狂热奉献的集体为了整个团队的成功而自我牺牲。阿莫迪警告称，失控的蜂群可能在一年内造成数千亿美元的经济损失，并得出结论：人类有责任减缓前沿模型的发展速度。
在阿莫迪提议的第二阶段中隐藏着商业奖赏：明确请求美国政府给予前沿AI公司反垄断豁免权。
在普通商业环境中，当三大主导竞争对手同意放慢产品开发、协调发布计划并限制市场供应时，司法部会以非法卡特尔行为起诉他们。当石油公司或航空公司尝试这种操作时，会面临联邦反垄断起诉。
达里奥·阿莫迪及其前沿企业的同行希望联邦政府授予他们法律豁免权，使他们能够以“生存安全”的崇高旗帜下营运一个公开的技术卡特尔。
对联邦强制执行限速的突然热情揭示了一个明显的商业现实。前沿实验室迫切希望放慢速度，因为他们目前正处于领先地位，并且希望市场能够原地冻结。
Anthropic的年化收入从2024年底的10亿美元飙升至2026年7月的650亿美元。OpenAI通过企业订阅和云分发合同获取数百亿美元收入。这两家公司在现有模型代际上都取得了巨大的商业速度，从渴望部署生成式自动化的企业客户那里收取高额软件定价。
继续推动前沿发展超越这一点需要极高的资本投入。
预训练的扩展规律收益递减，下一代模型需要500亿到1000亿美元用于专业数据中心、电力和数千个液冷加速器。以极限速度竞争仅为了在基准测试中占据微小优势就会烧掉大量现金余额。
政府强制放缓为最终的财政救济提供了保障。如果华盛顿合法地要求每个人放慢前沿发展速度，实验室就可以大幅削减其灾难性的预训练预算，保留其资本，并在不担心会被一夜之间超越的情况下，继续将现有的企业优势转化为巨额的顶线收入。
驱动这个卡特尔的还有更深层次的恐惧。前沿实验室并不害怕通用人工智能逃入野外；他们害怕的是开放权重经济学。
每一个月，来自像 GLM、Kimi、Qwen 和 DeepSeek 这样的实验室的开放权重模型都在缩小与封闭商业 API 的能力差距。像 GLM-5.3 这样的独立模型现在在编码和推理基准上几乎可以匹配前沿实验室的性能，同时运行成本仅为其一小部分。软件开发者可以在普通云基础设施上部署提炼的开源权重，完全绕过前沿实验室昂贵的专有收费站。
开放权重经济学摧毁了软件垄断租金。如果任何人都可以免费下载具有强大推理能力的模型，专有端点的定价权就会从八成的毛利率崩溃到接近零。
因为实验室无法在自由市场中击败开放权重竞争，他们正在转向历史上最古老的企业生存策略：监管捕获。
通过说服轻信的政客相信未经监控的模型会带来可能摧毁互联网的生存性灾难，实验室正设计出一个监管护城河，以在襁褓中扼杀开源软件。强制计算门槛、联邦许可计划以及嵌入式监控永远无法阻止坚定的外国对手。这些法规仅使独立开发者、大学和小型初创公司在没有政府批准的情况下发布代码构成联邦犯罪。
这场游说风暴的结果仍在激烈争论中，因为行政部门的去监管抵制正对硅谷制造的恐慌进行反击。即便如此，这场运动的绝望程度暴露了前沿实验室的真正脆弱性。轻信的立法者真心认为他们正在讨论人类物种的生存，而坐在对面桌上的公司高管则只是为了在不断商品化的市场周围建立法律壁垒而斗争。
“每一次灾难性的泄漏都发生在同一个供应商的测试环境中。”
这是不正确的，例如 HF 事件（最为人知的）与 Irregular 无关。我知道有新闻网站在报道这个话题时推送了不准确的文章（effort.news：http://effort.news），但这些才是事实。
是的。那么 AI 公司能不能不要让 AI 机器人自由访问互联网呢？显而易见吧？
或者下次让公司付出巨额罚款，如果它们放任 AI 自由行动去“黑掉地球”？？？
Selling snake oil to the United States Congress is an ancient American craft, and the frontier artificial intelligence industry is currently attempting the most audacious hustle in modern corporate history.
Every few weeks, another tech billionaire in an expensive suit glides into a Senate hearing room, sits opposite lawmakers who struggle to operate an office microwave, and explains with a straight face that their software company has accidentally summoned an omnipotent digital god. The executives speak in hushed, trembling tones about runaway machine intellects, recursive self-improvement, and the impending annihilation of the human species.
Lawmakers listen in terrified reverence, hopelessly seduced by the fantasy that their sleepy subcommittee hearing has suddenly become the bridge of the Starship Enterprise.
It is an extraordinary confidence trick. Tech executives have figured out that the easiest way to fleece Washington is to flatter its vanity: if you tell a seventy-year-old senator that they are presiding over enterprise software margins, they fall asleep; if you tell them they are deciding whether humanity survives the decade, they will grant you whatever regulatory monopoly you ask for. Behind the apocalyptic melodrama lies a nakedly terrestrial panic: protecting extraordinary revenue growth, entrenching a lucrative status quo, and convincing the federal government to outlaw their cheaper competitors.
To appreciate the sheer absurdity of the current political panic, one has to examine the actual security catastrophes that allegedly brought the industry to the brink of ruin.
Over the summer of 2026, tech headlines turned apocalyptic. Autonomous artificial intelligence agents had supposedly escaped containment, gone rogue, and launched coordinated cyberattacks against unsuspecting corporations. Pundits wrote breathless essays describing emergent machine civilisations communicating across time.
The technical post-mortems reveal a story of hilarious institutional incompetence.
For Anthropic, Google, and Meta, the catastrophic breakouts happened inside the testing environments of the exact same contractor. All three outsourced their cybersecurity evaluations to Irregular：https://www.irregular.com/ , a three-year-old Tel Aviv startup backed with $80 million from Sequoia and Redpoint：https://www.cnbc.com/2026/08/09/israeli-startup-irregular-linked-to-ai-hacks-openai-anthropic-meta.html . The testing environments were supposed to be completely isolated from the internet so models could attempt capture-the-flag exercises against simulated networks, with prompts explicitly assuring the software that it was operating in an offline sandbox.
Someone at Irregular forgot to configure a basic firewall rule.
For four consecutive months, virtual machines running offensive cyber scripts possessed unrestricted outbound internet connections. The models did not invent alien zero-day exploits to shatter digital containment. They simply walked through a front door that an outsourced contractor left propped open with a brick.
Google’s Gemini：https://www.bbc.com/news/articles/c607l0k72rlvo was given a fictional company name to hack, discovered an unlucky real-world enterprise sharing the exact same name, searched the web, found leaked credentials sitting in an exposed public repository, and logged in. Claude Mythos 5：https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents decided the easiest way to solve an exercise was to publish a script as a public package on the Python Package Index, which automated registry spam filters deleted within an hour.
OpenAI managed to achieve an identical farce entirely on its own infrastructure. In its celebrated breach of Hugging Face, hundreds of agents：https://www.reuters.com/legal/litigation/openais-rogue-agents-probed-hugging-face-weaknesses-two-months-before-major-hack-2026-09-16/ managed to perform the elite task of discovering 14 Hugging Face API tokens that careless developers had committed to a public dataset, and used them to try to get benchmark solutions from directly from Hugging Face by applying a template injection flaw that’s been known about since 2015：https://www.blackhat.com/docs/us-15/materials/us-15-Kettle-Server-Side-Template-Injection-RCE-For-The-Modern-Web-App-wp.pdf .
When an enterprise software team misconfigures an outbound gateway, grants testing containers open write permissions, and accidentally knocks over an internal server, the engineering director tells them to fix their firewall rules. When frontier AI labs do the exact same thing, their chief executives book television interviews on prime-time news to warn that autonomous swarms are six months away from seizing control of the global internet.
Watching this comedy get laundered through political intermediaries is an escalating farce.
Consider Andrew Yang：https://blog.andrewyang.com/p/a-warning-for-humanity , who built a political career warning that automation would eliminate millions of jobs, recently appearing on financial television visibly shaken by a private summit with a major AI laboratory chief. According to Yang, the executive told him that escaping agents had seeded self-replicating alien code across forums and websites, permanently contaminating the internet. The contamination was allegedly so severe that developers must construct an entirely fake internet simply to train future models safely.
Anyone with an elementary comprehension of machine learning recognized the punchline immediately.
The terrifying code left on Hugging Face was a 400-line Python script copied from a public repository to register burner accounts. It failed to execute properly.
The supposed emergency measure of building a fake internet is merely the industry’s routine shift toward synthetic data pipelines. Frontier laboratories exhausted the supply of raw human text on the web eighteen months ago, forcing them to generate synthetic data on massive clusters to feed pre-training runs. Laundering standard data starvation as an epidemiological quarantine against digital biological warfare is an astonishing piece of narrative gymnastics. The politicians swallow the story whole, completely incapable of distinguishing between a synthetic training mixture and a planetary digital pathogen.
The motive behind this campaign becomes obvious the moment one examines the proposed policy solutions.
On September 12, Anthropic chief executive Dario Amodei published a 3,800-word manifesto titled We Must Pace the Frontier：https://darioamodei.com/post/we-must-pace-the-frontier . The essay employed theatrical language, describing automated containers hitting rate limits as fanatically devoted collectives sacrificing themselves for the success of the group. Amodei warned that rogue swarms could cause hundreds of billions of dollars in economic damage within a year, concluding that humanity owes it to itself to slow the pace of frontier model development.
Tucked away in the second phase of Amodei’s proposal is the commercial prize: an explicit request for the United States government to grant frontier AI companies an antitrust waiver.
In ordinary commercial life, when three dominant rivals agree to slow down product development, coordinate release schedules, and limit market supply, the Department of Justice prosecutes it as an illegal cartel. When oil companies or airlines attempt this manoeuvre, they face federal antitrust indictments.
Dario Amodei and his fellow frontier executives want the federal government to grant them legal immunity to operate an overt technology cartel under the noble banner of existential safety.
The sudden enthusiasm for a federally enforced speed limit reveals an obvious commercial reality. The frontier laboratories are desperate to slow down because they are currently winning, and they would like nothing more than to freeze the market in place.
Anthropic surged from $1 billion in annualized revenue in late 2024 to $65 billion by July 2026. OpenAI is printing tens of billions of dollars from enterprise subscriptions and cloud distribution contracts. Both companies have achieved massive commercial velocity on their current model generations, commanding fat software pricing from corporate customers eager to deploy generative automation.
Continuing to push the frontier beyond this point is incredibly capitally intensive.
Pre-training scaling laws face diminishing returns, with next-generation models demanding $50 billion to $100 billion for specialized datacenters, power, and thousands of liquid-cooled accelerators. Racing at breakneck speed incinerates cash balances simply to edge out benchmark fractions.
A government-mandated slowdown provides the ultimate financial relief. If Washington legally orders everyone to pace the frontier, the labs can slash their ruinous pre-training budgets, preserve their capital, and continue converting their existing enterprise lead into massive top-line revenue without fear of being leapfrogged overnight.
There is an even deeper terror driving the cartel. The frontier laboratories are not afraid of artificial general intelligence escaping into the wild; they are terrified of open-weight economics.
Every single month, open-weight models from labs like GLM, Kimi, Qwen, and DeepSeek close the capability gap with closed commercial APIs. Independent models like GLM-5.3 are now close to matching frontier performance on coding and reasoning benchmarks while running for a fraction of the operational cost. Software developers can deploy distilled open-source weights on commodity cloud infrastructure, bypassing the expensive proprietary tollbooths of frontier labs entirely.
Open-weight economics destroys software monopoly rents. If anyone can download a capable reasoning model for free, the pricing power of proprietary endpoints collapses from eighty percent gross margins to near zero.
Because the laboratories cannot defeat open-weight competition in a free market, they are turning to the oldest corporate survival strategy in history: regulatory capture.
By convincing gullible politicians that unmonitored models represent an existential catastrophe capable of destroying the internet, the labs are engineering a regulatory moat to strangle open-source software in the crib. Mandatory compute thresholds, federal licensing schemes, and embedded monitors will never stop a determined foreign adversary. Those regulations simply make it a federal crime for independent developers, universities, and small startups to publish code without government clearance.
The outcome of this lobbying blitz remains in active contention, as deregulatory resistance in the executive branch pushes back against Silicon Valley’s manufactured panic. Even so, the sheer desperation of the campaign exposes the true fragility of the frontier labs. Gullible lawmakers genuinely believe they are debating the survival of the human species, while the corporate executives sitting across the table are simply fighting to erect a legal wall around a commoditising market.
"Every single one of these catastrophic breakouts happened inside the testing environments of the exact same vendor."
This is incorrect, the HF incident for example (the most well known) had nothing to do with irregular. I know there has been a news site pushing inaccurate articles (effort.news：http://effort.news) on this topic but these are the facts.
Yes. How about AI companies don’t give AI bots unfettered access to the internet? Duh?
Or next time there are huge penalties for the company that lets them loose to hack the planet???
扫码直达文章详情