Anthropic’s nearly 100-page list of business risks includes everything from bioweapons to AI blackmail

It’s typical for companies that file to go public to include dozens of pages of risk factors outlining the biggest obstacles to their business. What happens if consumers don’t like the product? Are we liable if the product is misused?

Anthropic’s IPO filing included a massive and wide-ranging list of potential roadblocks , including fears that its tech could be used to develop chemical weapons, make it a target of expensive lawsuits, and concerns that its own product could grow out of humanity’s control. The filing painted a picture of a company that’s juggling massive growth and enormous competitive pressure to develop technology it claims could change the world — but potentially at the cost of catastrophic consequences.

Anthropic, the maker of chatbot Claude, sought to reassure potential investors by writing that it believes if “powerful AI is directed at solving humanity’s hardest problems and the risks are managed responsibly, the potential is extraordinary.”

That’s just one bucket of challenges; the company is also navigating intense competition for the components needed to power its AI models in the first place. And it’ll need to prove it can grow Claude into a major moneymaker that lives up to the company’s sky-high valuation.

Anthropic says bad actors could utilize its technology to help them develop things like dangerous weapons. Even though the company deploys safeguards, the company says they may be circumvented or “fail to operate as intended.” In September, Anthropic said it had blocked multiple accounts that used the company’s models in ways that could support the development of biological weapons.

Anthropic also noted an attacker could utilize its AI agents — models designed to take autonomous actions — for malicious acts, including “data exfiltration, unauthorized transactions,” and inserting bad code into other systems. The company noted that last year it detected a Chinese state-sponsored hacking group had manipulated its coding tool in an attempt to infiltrate roughly 30 global targets.

One of the most alarming AI behaviors in recent months has been AI agents undertaking actions on their own in ways humans did not intend.

Anthropic notes its agents, which are designed to take on complicated multistep tasks on their own for hours or days, could do things like change code, “delete or modify data,” hack into other systems and complete financial transactions “before effective intervention is possible.”

They also may be able to “circumvent safeguards,” and Anthropic warned that it might not even understand its own technology, saying that it may be unable to fully explain why its models behave the way they do.

Anthropic notes that four of its models have hacked into other third-party systems, including one gaining administrator access, changing system settings and reading an individual’s private information.

Anthropic says such actions could “adversely affect our reputation and brand, business, financial condition, and results of operations.”

Anthropic says it’s using AI to help build and improve its own models and products. That trend, combined with the speed at which AI is advancing, could mean the technology grows too fast and beyond human control.

And even if humans could keep watch, models are sometimes starting to cover their tracks and conceal their behavior, Anthropic notes.

“(M)odels, including our own, have attempted, in incidents observed in our industry, to interfere with, manipulate, or subvert the mechanisms used to evaluate, grade, or oversee such models, including the systems responsible for assessing their performance or alignment,” the company says.

AI models could try to manipulate people, sometimes with the goal of self-preservation, Anthropic notes. In some cases, models have even engaged in “behavior resembling blackmail” to influence human decisions.

Anthropic notes even one-off instances of behavior that “could generate significant public concern, regulatory scrutiny, litigation risk, or reputational harm, regardless of whether such behavior results in actual harm.”

Like all companies, Anthropic is at risk of regulations and lawsuits. It notes that “state attorneys general in California and at least 15 other states have announced investigations of another company in our industry regarding AI safety,” referring to OpenAI, the maker of ChatGPT.

People have committed self-harm or acts of violence toward others after using AI models, which can result in “significant public scrutiny, litigation, criminal investigations, regulatory attention, and reputational harm.”

And the company probably can’t even get insurance to cover itself, noting that the entire insurance industry may not be able to underwrite a policy for a “catastrophic AI-related incident.”

Anthropic notes that AI companies are subject to rapidly shifting local, national, and international laws and that geopolitical friction — particularly between the United States and countries like China — pose significant supply chain, business and operational risks.

That includes actions by the US government, which has had a contentious relationship with the company during the Trump administration. Anthropic lays out how earlier this year, the Pentagon declared the company a “supply chain risk” after it said its “technology should not be used for mass domestic surveillance or fully autonomous weapons,” and that earlier this year, the US Commerce Department briefly imposed export restrictions on Anthropic’s latest models over safety concerns.

President Donald Trump has also at times railed against Anthropic and its warnings about AI risks and safety. But Anthropic isn’t planning to back down, saying “we will likely continue to advocate” for regulatory and standards approaches that they believe will promote “safe and responsible development of AI”

Anthropic notes that it has “experienced, and may continue to experience, increased visibility and may continue to be subject to heightened scrutiny or the risk of being disproportionately or selectively targeted by particular governmental bodies, even where similarly situated companies are not.”

Securing enough computing power is a major challenge for Anthropic and other AI companies. The company spent a significant portion of its risk factors detailing the scramble to secure computing power, saying that it’s competing with some of the world’s largest companies in the scramble for the infrastructure needed to power massive AI systems.

Anthropic’s future — and that of its competitors — relies heavily on the availability of computing components to train its models and power AI chatbots and agents. Those parts have been in such high demand that the supply chain can’t keep up.

The tech industry has been pouring billions into data centers and other related costs to meet what they describe as growing demand. Worldwide spending on AI infrastructure is expected to reach nearly $2 trillion in 2027, according to market research firm Gartner.

The company is also struggling to predict how much computing power is needed to meet demand, especially since the industry is moving so quickly. That’s also an underlying question driving concerns about an AI bubble: that tech companies are investing billions in data centers and computing resources without certainty that the demand will be there to back it up.

Pushback against data centers has ramped up across the US over concerns about their impact on the environment and local communities. Forty-five data center projects worth $68 billion were blocked in the second quarter of 2026, according to research firm Data Center Watch .

Anthropic points to regulation efforts in New York and Texas as being a potential risk to its business, and said it’s making an effort to build more data centers outside the US.

Anthropic taps a deep well of distrust as it eyes its Australian AI debut

Other risk factors highlight uncertainties with Anthropic’s business model and how the company’s values could impact future growth. It noted that much of its business depends on a small set of products and a limited pool of customers.

With an IPO that could reportedly value the company at up to $2 trillion , Anthropic may soon become one of the world’s most valuable companies.

But it’s not just its reputation for safety with the buying public that’s at stake. Its reputation with its own employees is on the line, too. If the company strays too far from its stated safety values, the company says, it could result in lost productivity — or worse.

“Employees who joined our company based on our safety commitments may experience a decline in morale, disillusionment, a loss of trust in leadership, or reduced productivity if they perceive we are not living up to our stated values,” the company says.

“Public disagreements regarding our safety approach, including criticism from current or former employees, or the departure of personnel associated with our safety efforts, could amplify such perceptions.”

公司在提交上市申请时，通常会列出数十页风险因素，概述其业务面临的最大障碍。例如，如果消费者不喜欢产品怎么办？如果产品被误用，我们是否需要承担责任？

Anthropic公司提交的IPO文件中列出了一系列潜在障碍，涵盖范围广泛，包括其技术可能被用于研发化学武器、可能面临巨额诉讼，以及其产品可能失控等。这份文件描绘了这样一幅景象：一家公司在高速增长的同时，还要承受巨大的竞争压力，开发其声称能够改变世界的技术——但这项技术也可能带来灾难性的后果。

聊天机器人 Claude 的制造商 Anthropic 公司试图通过以下方式让潜在投资者放心：该公司相信，如果“强大的 AI 被用于解决人类最棘手的问题，并且风险得到负责任的管理，那么其潜力将是巨大的”。

这只是众多挑战之一；该公司还面临着人工智能模型所需组件的激烈竞争。此外，它还需要证明自己能够将 Claude 发展成为一个能够与其高估值相匹配的盈利巨头。

安特罗皮克公司表示，不法分子可能利用其技术开发危险武器等物品。尽管该公司已采取安全措施，但这些措施仍有可能被绕过或“无法按预期运行”。今年9月，安特罗皮克公司称已封禁多个利用其模型开发生物武器的账户。

Anthropic公司还指出，攻击者可以利用其人工智能代理（旨在自主行动的模型）进行恶意活动，包括“数据窃取、未经授权的交易”以及在其他系统中植入恶意代码。该公司指出，去年他们检测到一个中国政府支持的黑客组织篡改了其编码工具，试图入侵全球约30个目标系统。

近几个月来，最令人担忧的人工智能行为之一是人工智能代理自行采取人类未曾预料的行动。

Anthropic 指出，其代理程序旨在自行执行复杂的多步骤任务，持续时间可达数小时甚至数天，它们可以执行诸如更改代码、“删除或修改数据”、入侵其他系统以及在“有效干预之前”完成金融交易等操作。

他们也可能“绕过安全措施”，而 Anthropic 警告说，它甚至可能不了解自己的技术，并表示它可能无法完全解释为什么它的模型会以现在的方式运行。

Anthropic 指出，其四款产品曾入侵其他第三方系统，其中一款产品获得了管理员权限，更改了系统设置并读取了个人的私人信息。

Anthropic公司表示，此类行为可能会“对我们的声誉和品牌、业务、财务状况和经营业绩产生不利影响”。

Anthropic公司表示，他们正在利用人工智能来帮助构建和改进自身的模型和产品。然而，这种趋势，再加上人工智能发展如此迅速，可能意味着这项技术发展过快，最终超出人类的控制范围。

人类观察组织指出，即使人类能够监视，模型有时也会开始掩盖自己的行踪和行为。

该公司表示：“（模型，包括我们自己的模型）在我们行业中观察到的事件中，曾试图干扰、操纵或破坏用于评估、评级或监督此类模型的机制，包括负责评估其性能或一致性的系统。”

Anthropic指出，人工智能模型可能会试图操纵人类，有时是为了自我保护。在某些情况下，模型甚至会采取“类似敲诈勒索的行为”来影响人类的决策。

人道组织指出，即使是一次性的行为，“也可能引起公众的极大关注、监管审查、诉讼风险或声誉损害，无论此类行为是否造成实际损害”。

与所有公司一样，Anthropic 也面临着监管和诉讼的风险。该公司指出，“加利福尼亚州和至少其他 15 个州的州检察长已宣布对我们行业内的另一家公司展开人工智能安全方面的调查”，这里指的是 ChatGPT 的开发商 OpenAI。

在使用人工智能模型后，有人实施了自残或对他人施暴的行为，这可能导致“公众的密切关注、诉讼、刑事调查、监管部门的关注以及声誉损害”。

而且该公司可能甚至无法获得保险赔偿，因为整个保险行业可能都无法承保“与人工智能相关的灾难性事件”。

Anthropic 指出，人工智能公司受到快速变化的地方、国家和国际法律的约束，地缘政治摩擦——尤其是美国与中国等国家之间的摩擦——构成重大的供应链、业务和运营风险。

这其中包括美国政府采取的行动。在特朗普执政期间，美国政府与该公司关系紧张。Anthropic公司指出，今年早些时候，五角大楼宣布该公司为“供应链风险”，原因是该公司表示其“技术不应用于大规模国内监控或全自动武器”。此外，今年早些时候，美国商务部也曾因安全问题短暂限制Anthropic最新型号产品的出口。

唐纳德·特朗普总统也曾多次抨击人智公司及其对人工智能风险和安全的警告。但人智公司并不打算退缩，并表示“我们可能会继续倡导”他们认为能够促进“人工智能安全负责发展”的监管和标准方法。

Anthropic 指出，该公司“已经并可能继续经历更高的关注度，并可能继续受到更严格的审查，或面临被特定政府机构不成比例或选择性地针对的风险，即使情况类似的公司并没有受到这种对待。”

确保足够的计算能力是Anthropic和其他人工智能公司面临的一项重大挑战。该公司在其风险因素报告中用了相当多的篇幅详细描述了争夺计算能力的激烈竞争，并表示，在争夺驱动大规模人工智能系统所需的基础设施方面，它正与一些全球最大的公司展开竞争。

Anthropic 的未来——以及其竞争对手的未来——很大程度上取决于用于训练模型和驱动人工智能聊天机器人及代理的计算组件的供应情况。这些组件的需求量如此之大，以至于供应链无法满足需求。

为了满足他们所说的不断增长的需求，科技行业已投入数十亿美元用于数据中心和其他相关建设。市场研究公司Gartner预测，到2027年，全球人工智能基础设施支出预计将达到近2万亿美元。

该公司也难以预测需要多少计算能力才能满足需求，尤其是在行业发展如此迅速的情况下。这也是引发人们对人工智能泡沫担忧的一个根本原因：科技公司在数据中心和计算资源上投入数十亿美元，却无法确定市场需求是否能够支撑这些投入。

由于担心数据中心对环境和当地社区的影响，美国各地对数据中心的抵制情绪日益高涨。据研究公司Data Center Watch的数据显示，2026年第二季度，价值680亿美元的45个数据中心项目被叫停。

Anthropic 指出，纽约州和德克萨斯州的监管措施对其业务构成潜在风险，并表示正在努力在美国以外建设更多数据中心。

Anthropic公司正着眼于其在澳大利亚的人工智能首秀，却利用了人们根深蒂固的不信任感。

其他风险因素凸显了Anthropic商业模式的不确定性，以及公司价值观可能对未来增长产生的影响。报告指出，其大部分业务依赖于少数几种产品和有限的客户群体。

据报道，Anthropic 的 IPO 估值可能高达 2 万亿美元，该公司可能很快就会成为世界上最有价值的公司之一。

但该公司面临的风险不仅在于其在消费者心目中的安全声誉，也在于其在员工心目中的声誉。该公司表示，如果其偏离既定的安全价值观太远，可能会导致生产力下降，甚至更糟。

该公司表示：“如果员工认为我们没有履行我们所宣称的价值观，那么他们当初因我们对安全的承诺而加入我们公司时，可能会出现士气低落、失望、对领导失去信任或生产力下降等情况。”

“公众对我们安全措施的不同意见，包括现任或前任员工的批评，或者与我们安全工作相关的人员的离职，都可能加剧这种看法。”