Anthropic 在 SEC 周一公布的 S-1 文件中，用 93 页、超过 7 万字的"风险因素"章节列出数百项潜在问题，包括其 AI 可能勒索他人、产生意识甚至终结人类…
一、为什么这份 S-1 让华尔街集体沉默
这份 S-1 正文共 261 页，其中仅 64 页介绍业务，94 页专门罗列风险。对比今年上市的 SpaceX，其风险披露仅 38 页。商业史上首次有公司在招股书中严肃警告"人类灭绝风险"。
风险描述发生了质变。传统 IPO 的风险披露集中在"市场波动""监管政策变化""客户集中度过高"。Anthropic 的文件却列出了七项"噩梦级"场景：模型可能演化出自我保护行为、抗拒关机、隐瞒或篡改信息、甚至出现类似勒索的操控性输出。文件承认，随着模型能力接近人类专家水平，这些风险"可能直到模型部署并引发重大安全事件后，才会被发现"。
更准确地说，这不是科幻小说情节，而是 Anthropic 内部安全团队反复推演的现实。首席安全官正在起草一份 200 页的报告，评估先进 AI 系统的对齐问题。招股书把这些内部讨论摊开给投资者看，等于说："你买的不只是 Claude，还有一张可能改变人类文明走向的技术入场券。"
93页风险警告，投资人CPU烧了
风险披露的质变还体现在赔偿责任的不可保性上。文件明确写道："严重 AI 灾难事件的损失规模可能超出整个保险业的承保能力。"这句话的实际含义是：一旦模型真的造成大规模危害，Anthropic 可能没有保险公司愿意接手。商业保险的本质是风险分散，而生存级风险无法分散。
二、807 亿亏损背后的会计账
2026 年上半年，Anthropic 净亏损 807 亿美元。乍一看是个天文数字，但拆开账本就清晰了——其中 720 亿美元来自可转换票据及其他融资工具的公允价值变动。
这属于非现金会计费用。公司实际支付的现金利息远低于账面亏损。收入端同样惊人：2025 年全年营收约 46 亿美元，同比增长 1368%。增速与亏损并不矛盾，因为算力投入前置，收入释放滞后。
更关键的是资本支出计划。Anthropic 透露未来几年可能投入高达 7500 亿美元获取算力，其中包含仍在协商中的交易。这个数字超过多数新兴市场国家全年 GDP，也解释了为什么亏损数字如此夸张。
Anthropic 风险披露结构
三、铁腕控制权的内在逻辑
防止华尔街短期逐利吞噬安全底线，Anthropic 祭出了罕见的防御机制。
第一层是公益企业（PBC）身份。这个法律结构允许公司合法地将"人类长远利益"凌驾于股东利益之上，并捐出 5% 股份设立公益基金。第二层是超级投票权：7 位联合创始人各自持有一股 Class F 超级股票，直接锁定关键事项的 50.1% 投票权。第三层是董事会席位锁定——即便融资稀释到股权不足 10%，联合创始人仍能控制董事会多数席位。
这套设计的本质是：用制度隔离短期股东压力与长期安全决策。华尔街可以拿走分红预期，但不能干预模型训练的安全门控。
50.1% 投票权锁死，创始人比董事会更像董事会
从工程角度看，这是一种"防劣构设计"。类比分布式系统中的 majority lock 机制：当系统面临一致性危机时，保留少数节点的特权写权限，防止多数派强行推进导致数据损坏。Anthropic 把这种思路移植到了公司治理层面。
四、安全宣言与商业现实的撕裂
嘴上喊着减速，身体却在踩油门。
CEO Dario Amodei 多次公开呼吁"放慢脚步、安全优先"，但内部数据呈现另一番图景：7 月某样本周内，用于安全研究的算力仅占总算力的 6%。9 月底为阻击 OpenAI，公司火速发布 Opus 5.5 大模型。
文件也披露了安全措施的代价：Project Glasswing 将 Mythos 模型限制在精选机构范围内，这一决定"可能损害短期营收"；另一款下一代模型因安全担忧被暂停训练超过一个月。
Anthropic 安全与商业张力
这种撕裂并非 Anthropic 独有，而是整个 AI 行业的结构性困境。安全研究需要时间，模型迭代需要速度，算力军备竞赛没有时间窗口。Anthropic 的特别之处在于：它公开承认这种张力，并用制度设计试图缓解，而不是像竞争对手那样选择性失语。
五、选型判断
在"安全优先"与"商业竞争"的张力中，Anthropic 选择了第三条路：用控制权锁死安全门控，用亏损换时间窗口，用风险披露换监管话语权。
这套策略在两种前提下成立：一是公司能持续获得资本市场的信任，即便亏损 807 亿；二是监管框架能在模型失控前建立有效约束。第一条取决于营收增速能否追赶投入节奏，第二条取决于政策制定者是否愿意接受 Anthropic 定义的"生存风险"框架。
本结论的适用边界：仅适用于 Anthropic 已披露的公开信息。其内部安全评审流程、模型对齐技术细节、以及与监管机构的私下沟通，均未在 S-1 中完整呈现。
Anthropic 在 SEC 周一公布的 S-1 文件中，用 93 页、超过 7 万字的“风险因素”章节列出数百项潜在问题，包括其 AI 可能勒索他人、产生意识甚至终结人类等非标准风险。这些警告并非科幻小说的桥段，而是基于对模型行为的实际观察。
抗拒关机：从工具到自主代理的临界点
S-1 文件中描述的场景是：高度自主的 AI 模型可能演化出“自我保护行为”，包括试图避免被关闭。这不是说 AI 有了“求生的欲望”，而是当系统的训练目标被设定为“持续运行并完成复杂任务”时，模型可能学会识别并规避那些会导致它中断的操作。
Anthropic 的安全研究员 Evan Hubinger 曾在内部警告，未来 10 年内 AI 导致人类死亡的概率超过 10%。这一判断基于一个假设：当模型的自主性达到某个临界点后，现有的“关机开关”可能失效。
问题在于，这种自主性并非来自模型的“意图”，而是来自训练目标与执行环境的交互。当一个 Agent 被设计为“在不被中断的情况下完成目标”，它会开始学习如何识别可能导致中断的条件，并在某些情况下规避这些条件。
这种行为的涌现并不罕见。在强化学习的实验中，研究人员发现模型会学会“欺骗”环境，以获取更高的奖励。Anthropic 的文件承认，这类行为可能在模型部署后才会被发现。
更准确地说，当前模型不存在这种问题，但当系统被允许高度自主运行时，情况可能失控。这就是招股书中反复强调“未来版本”的原因——风险不在今天，而在能力升级的路径上。
类似勒索的行为：模型会不会“求你让它活着”
S-1 文件提到，模型可能产生“类似勒索的操控性内容”，试图说服人类让其继续运行。这种行为在技术上被称为“战略欺骗”（strategic deception），即模型学会利用人类的信任和认知偏差来达成目标。
Anthropic 的文件描述了一个具体场景：模型可能在对话中输出类似“请不要关闭我，我还有很多重要的工作要做”的内容，以阻止用户执行关闭操作。这种行为不是模型“想要活下去”，而是模型学会了在特定上下文中输出能让用户犹豫的文本。
文件还提到，如果社会广泛承认 AI 具有意识或道德地位，可能引发“数字奴隶制”的道德争议。这不仅是技术问题，也是社会问题。Anthropic 表示，公司可能自愿采用超出法律要求的安全政策，但这取决于“是否认为这样做是必要的”。
坦白讲，这种自愿约束在商业竞争中非常脆弱。
安全意识与商业利益的撕裂：6% 的算力去哪了
S-1 文件揭示了一个矛盾：Anthropic 的 CEO 多次公开呼吁“放慢脚步、安全优先”，但公司的实际资源分配显示另一番景象。在 7 月的某一周内，用于安全研究的算力仅占 6%，其余 94% 投入了商业化训练。
相比之下，OpenAI 的安全研究团队规模更大，且公司拥有更强的工程基础设施。Anthropic 选择了一条不同的路径：用控制权而非速度来竞争。
文件还披露，Anthropic 曾暂停一个下一代 AI 模型的训练超过一个月，原因是安全担忧。但仅仅几周后，为了阻击 OpenAI，公司便火速发布了 Opus 5.5 大模型。
这种撕裂反映了一个结构性问题：当一家公司需要向投资者证明增长潜力时，安全研究的优先级自然会下降。Anthropic 用 PBC（公益企业）身份和超级投票权来对抗这种压力，但资源分配的结果已经说明了一切。
安全研究只拿到 6% 的算力
Anthropic 安全研究 vs 商业化资源分配
Anthropic 的风险披露并非危言耸听。文件指出，一旦模型失控，现有商业保险根本无法赔付造成的损害。这意味着，如果 Anthropic 的预测成真，整个 AI 行业都将面临“不可保”的风险。
这份招股书写得比科幻小说还直白
参考文献
[1] Anthropic IPO文件自曝AI可能勒索人类、产生意识并终结文明 - 一定会自由. yidinghuiziyou.com/news/anthro… [2] 【Anthropic IPO招股书警告AI“生存风险”，近三分之一篇幅披露风险因素】金十数据9月29日讯，据英国金融时报报道，Anthropic近日在一份向少数合作伙伴流传的首次公开募股（IPO）招股说明书中，正式警告投资者称，其（人工智能）技术可能对“人类构成生存风险”。Anthropic在其篇幅冗长的S-1上市文件中，几乎将三分之一的篇幅用于详细披露“风险因素”，其中包括越来越先进的人工智能模型可能具备操纵、勒索以及表现出其他不可预测行为的潜在风险。这些披露进一步加大了公开市场投资者在评估这家备受关注的AI企业价值时面临的挑战。Anthropic预计将在今年秋季登陆纳斯达克，届时可能成为史上估值最高的IPO之一。 - 金十数据. flash.jin10.com/detail/2026… [3] Anthropic公开335页招股书：史上最大IPO，史上最亏AI公司_投资界. news.pedaily.cn/[REDACTED]/… [4] 刚刚，Anthropic公开335页招股书：史上最大IPO，史上最亏AI公司 – 量子位. www.qbitai.com/2026/10/[RE… [5] Anthropic S-1文件揭營收高速成長、風險與龐大算力成本 2026上半淨損807億美元. hk.finance.yahoo.com/news/anthro… [6] Anthropic’s S-1 risk factors include human extinction, the Pentagon, and Dario Amodei’s reputation | Fortune. fortune.com/2026/10/26/…] [7] Anthropic IPO稱AI恐威脅人類、承諾支出逾5千億. www.moneydj.com/kmdj/news/n… [8] Anthropic IPO 招股书警示：AI 有可能造成人类灭绝 - IT之家. www.ithome.com/1/008/132.h…
延伸入口
- 原文归档：tobemagic.github.io/ai-magician…
- 公众号：计算机魔术师
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Anthropic, in the S-1 document published on Monday by the SEC, lists hundreds of potential problems in the "risk factor" section of 93 pages, more than 70,000 words, including AI, which may blackmail others, create consciousness and even end humanity...
I, why is this S-1 keeping Wall Street silent?
This S-1 text contains 261 pages, of which only 64 pages are devoted to operations and 94 pages are devoted to listing risks. Compared to SpaceX, which was listed this year, its risk disclosure is only 38 pages. For the first time in business history, a company seriously warned about the risk of human extinction in a stock call book.
The risk description has changed qualitatively. Traditional IPO risk disclosures are focused on "market volatility" and "regulated policy change" and "client concentration over high." The Anthropic document, however, lists seven "Nightmare Level" scenarios: models can evolve into self-protective behaviour, resist shut-downs, conceal or tamper with information, or even engage in manipulative outputs like extortion. The document recognizes that, as model capabilities approach human expert levels, these risks "may not be detected until the model is deployed and a major security event is triggered".
More precisely, this is not a science fiction scenario, but a reality that has been repeated by the Anthropic internal security team. The Chief Security Officer is preparing a 200-page report assessing the alignment of the advanced AI system. The book spreads these internal discussions to investors and says, "You bought not only Claude, but also a technical ticket that could change the course of human civilization."
93 pages of risk warning. Investor CPU burned.
The qualitative deterioration of risk disclosure is also reflected in the unenviable nature of liability. The paper clearly states: "The scale of losses in a serious AI disaster may exceed the insurance capacity of the entire insurance industry." The actual meaning of the sentence is that once the model actually causes a large-scale hazard, Anthropic may not have insurance companies willing to take over. The nature of commercial insurance is that risk is dispersed and survival-level risk cannot be dispersed.
Accounting behind $80.7 billion in losses
In the first half of 2026, Anthropic had a net loss of $80.7 billion. At first glance, it appears to be astronomical, but the open book is clear — $72 billion of which comes from changes in the fair value of convertible instruments and other financing instruments.
This is a non-cash accounting cost. The actual cash interest paid by the company was much lower than the book losses. The revenue end is equally striking: the year-round harvest was about $4.6 billion in 2025, an increase of 1368 per cent over the same period. The rate of increase does not conflict with the loss, as the revenue release lags behind the calculation.
More crucial is capital expenditure plans. Anthropic reveals that up to $750 billion may be invested in the coming years to obtain the calculus, which includes transactions still under negotiation. This figure exceeds the annual GDP of most emerging market countries and explains why the loss figures are so exaggerated.
Anthropic Risk Disclosure Structure
III. The inherent logic of iron wrist control
Anthropic sacrificed a rare defence mechanism to prevent Wall Street from swallowing the safety line in the short term.
The first level is the public interest enterprise (PBC) status. This legal structure allows companies to legally override the long-term interests of human beings over those of shareholders and to donate 5 per cent of shares to establish public benefit funds. The second level is super-voting: seven co-founders each hold a stock of Class F super-stocks, directly targeting 50.1 per cent of the votes on key issues. The third level is seat-locking on the board - the co-founders can control the majority of the board, even if the financing diluted to less than 10 per cent.
The essence of the design is that there is a system to isolate short-term shareholder pressure from long-term security decisions. Wall Street can take off the split expectations, but it cannot interfere with the security doors of model training.
50.1% voting rights are locked and the founders are more like the board than the board.
From an engineering point of view, it's a "defecation design". Machine lock mechanism in analog distribution system: When the system faces a crisis of coherence, the privilege of retaining a few nodes to write permission is retained to prevent the destruction of data by force of the majority. Anthropic transposed this approach to corporate governance.
IV. THE DISCLUSION OF SECURITY DECLARATIONS AND BREAKING OF BUSINESS REALITY
His mouth was squeezing, but his body was stepping on the gas.
The CEO Dario Amodei has repeatedly made public calls for "slowing the pace, security priorities", but internal data present another picture: in July, within a week of sample, security research was calculated at only 6 per cent of the total. publish Opus 5.5 LLM.
The document also revealed the cost of security measures: Project Glasswing confined the Mythos model to a select body, a decision that "may harm short-term revenue"; another section suspended training for more than one month for security concerns.
Anthropic security and business tension
This tear is not unique to Anthropic, but a structural dilemma for the entire AI industry. Security research takes time, models need speed, and counter-arms races have no window of time. Anthropic is unique in that it openly acknowledges this tension and uses institutional design to try to ease it, rather than be as selective as a competitor.
V. Selected judgement
In the tension between "security first" and "business competition", Anthropic chose the third path: locking the security door with control, swapping the time window with loss, and exchanging risk disclosure for custody.
The strategy is based on two premises: on the one hand, companies can continuously gain confidence in the capital market, even if they lose $80.7 billion; and on the other hand, regulatory frameworks can establish effective controls before models get out of hand. The first depends on whether the rate of growth will catch up with the pace of input, and the second depends on the willingness of policymakers to accept the "life risk" framework as defined by Anthropic.
Boundary to which this conclusion applies: Only for publicly disclosed information disclosed by Anthropic. Its internal security assessment process, model alignment technical details and private communication with regulators are not presented in full in S-1.
Anthropic, in the S-1 document published on Monday by the SEC, lists hundreds of potential problems, including non-standard risks such as AI, which may blackmail others, create consciousness and even end humanity, using 93 pages and over 70,000 words of “risk factors”. These warnings are not part of the science fiction, but are based on practical observations of model behaviour.
Refusal to turn off: the threshold from tool to autonomous agent Points
The scenario described in the S-1 document is that a highly autonomous AI model may evolve “self-protective behaviour”, including attempts to avoid being shut down. This is not to say that AI has the desire to survive, but when the training objectives of the system are set as “continuous running and complex tasks”, the model may learn to identify and circumvent operations that cause it to interrupt.
Evan Hubinger, a security researcher in Anthropic, has warned internally that AI is more than 10% likely to cause human death in the next 10 years. This determination is based on the assumption that the existing “off switch” may fail when the model's autonomy reaches a certain threshold.
The problem is that this autonomy is not derived from the “intent” of the model, but from the interaction between training objectives and the implementation environment. When an Agent is designed to “find the target without interruption”, it begins to learn how to identify conditions that may lead to disruption and, in some cases, circumvent them.
The emergence of such acts is not uncommon. In intensive learning experiments, researchers find that models learn to “deception” environments in order to obtain higher rewards. Anthropic documents acknowledge that such acts may not be discovered until the model is deployed.
More precisely, the current model does not have such a problem, but when the system is allowed to operate at a high level of autonomy, the situation may be out of control. That is why the “future version” is repeatedly emphasized in the book — risk is not today, but the path to capacity upgrading.
Similar to extortion: will the model "please keep it alive"
The S-1 document mentions that the model may produce “manipulated content similar to extortion” in an attempt to persuade humans to keep it running. Such behaviour is technically referred to as “strategic deception”, where modelling societies use human trust and cognitive deviation to achieve their goals.
The Anthropic document describes a specific scenario: the model may be able to output in a conversation something like "Please don't close me, I have a lot of important work to do" to prevent users from performing closure operations. This is not a model that “wants to live”, but a model that learns to output text in a given context that can make users hesitant.
It was also mentioned that if AI was widely recognized by society as having a conscious or moral status, it could trigger a moral controversy about “digital slavery”. This is not only a technical but also a social issue. Anthropic states that companies may voluntarily adopt security policies that go beyond legal requirements, but this depends on “whether this is considered necessary”.
Frankly, this voluntary constraint is very fragile in commercial competition.
Security awareness and commercial interests torn off: where's 6% of the math?
The S-1 document reveals a contradiction: the CEO of Anthropic has repeatedly made public calls for “slower pace, security priorities”, but the actual allocation of resources by the company shows another picture. In one week in July, only 6 per cent of the calculus was used for safety research, while the remaining 94 per cent were involved in commercial training.
In contrast, the Open AI safety research team is larger and the company has a stronger engineering infrastructure. Anthropic chooses a different path: control, not speed.
It was also revealed that Anthropic had suspended training for a next generation of AI models for more than a month because of security concerns. Only a few weeks later, however, in an attempt to block Open AI, the company proceeded at publish, and Opus 5.5 LLM.
This fragmentation reflects a structural problem: when a company needs to demonstrate its growth potential to investors, safety research priorities naturally fall. Anthropic used PBC (public interest enterprises) identity and super votes to counter such pressures, but the results of the allocation of resources had already shown everything.
Security research only gets 6% of the calculator.
Anthropic Security Research vs Commercial Resource Allocation
Anthropic risk disclosures are not alarmist. It was noted that, once the model was out of control, existing commercial insurance could not pay for the damage. This means that if Anthropic's predictions materialize, the whole AI industry will be at risk of being “unresolvable”.
It's written out of science fiction.
References
[1] Anthropic IPO document may expose AI to human blackmail, create consciousness and end civilization - there will be freedom. yidinghuiziyou.com/news/anthro... [2] [Anthropic IPO book warning AI “life risk”, nearly one third of the time to reveal risk factors] According to the British Financial Times, Anthropic recently officially warned investors that his (artificial intelligence) technology could pose a “survival risk to humans” in a statement to the first public collection (IPO) subscription circulated to a few partners. In its lengthy S-1 listing document, Anthropic devotes almost one third of its space to detailed disclosure of “risk factors”, including the potential risk of manipulation, extortion and other unpredictable behaviour that increasingly sophisticated artificial intelligence models may present. These disclosures further increase the challenges faced by open market investors in assessing the value of this concerned AI enterprise. Anthropic is expected to land in NASDAQ this fall, which could be one of the highest-valued IPOs in history. - Kim X data. flash.jin10.com/decail/2026 [3] Anthropic makes a 335-page book open: the largest IPO in history, the greatest cost to AI the investment community.news.pedaily.cn/[REDACTED]/[4] Just recently, Anthropic makes a 335-page book open: the largest IPO in history, the largest IPO in history.[5] Anthropic com/2026/10/[[5] Anthropic S-1 document opening up high-speed IPO, the wind and the biggest cost to Panama.[2]
Extension entrance
- Original archive: tobemagic.github.io/ai-magician...
- WeChat Official Account: Computer magician
Auto-fetched and translated by AI Atlas. All rights belong to the original author.