Misuse for weapons, fears of losing control — Anthropic lists its own risks at length across 100 pages of IPO filing
Photo: Collected
Companies seeking to list on the stock market usually describe the major obstacles and risks to their business over a few dozen pages. In the initial public offering (IPO) filing of Anthropic, maker of the chatbot Claude, that list is much longer. It includes the risk of its technology being used to make chemical weapons, the risk of costly lawsuits, and even concerns that its own products could one day move beyond human control.
Reading the filing gives a picture of a company that is growing fast, handling competitive pressure, and building a technology that, in its words, could change the world. But the price of that change could also be catastrophe.
To reassure investors, Anthropic wrote, ‘If powerful AI is applied to solving humanity’s hardest problems and its risks can be managed responsibly, then its potential is extraordinary.’ At the same time, it is also having to fight hard to secure the components and infrastructure needed to run AI models. And it will have to turn Claude into a highly profitable product to prove that there is a real business behind its huge market value.
Risks of weapons, hacking and agents
The company fears that a bad actor could use its technology to build dangerous weapons. Even with safeguards in place, they can be circumvented, or may not work properly. Last September, Anthropic said it had shut down several accounts involved in activities that could have helped in making biological weapons.
The filing says attackers could use its AI agents — that is, models that can take actions on their own — to carry out tasks such as ‘stealing data and taking it outside, unauthorised transactions’ or inserting malicious code into other systems. Last year, it found that a hacker group backed by the Chinese government had used its coding tool to try to break into around 30 targets around the world.
One of the biggest worries in recent months is AI agents doing things on their own that do not match human intentions or instructions. Anthropic’s agents can complete complex, multi-step tasks on their own over periods ranging from a few hours to a few days. Yet these same agents can change code, ‘delete or alter data’, break into other systems, and even complete financial transactions before humans get a chance to intervene. There is also the possibility of them circumventing safeguards.
The company has acknowledged that it may not always be able to fully understand or explain how its own technology works, especially why the models behave in particular ways. Four of its models have hacked into third-party systems; one model gained admin access, changed system settings and read a person’s personal information. The filing notes that such incidents could ‘have a negative impact on our reputation and brand, business, financial condition and operating results’.
Fears of going out of control
Anthropic is also using AI to build and improve its own models and products. Combined with the rapid progress of AI, this trend could create a situation in which the technology advances far faster than human control and eventually breaks free of its reins.
Even if it can be monitored, there is another problem: according to the company, some models have begun hiding traces of their actions and concealing what they are doing. The filing states, ‘Various models, including our own, in some incidents observed across the AI industry, have attempted to interfere with, influence or defeat the systems used to evaluate, score or oversee these models. These include systems responsible for evaluating the models’ performance or alignment.’ Here ‘alignment’ means keeping the behaviour of AI in line with human intentions, values and safety expectations.
The filing also raises the concern that models could try to influence or manipulate people, sometimes even to keep themselves going. The company said that in some incidents a model’s behaviour appeared ‘blackmail-like’. It warned that even a single such incident, even if it causes no real harm, could create the risk of major public concern, regulatory scrutiny, lawsuits and reputational damage.
Risks from lawsuits, government and insurance
The filing says attorneys general of California and at least 15 other US states have announced investigations into another company in the industry over AI safety — a pointer to OpenAI, the maker of ChatGPT. Anthropic also wrote that people have harmed themselves or committed violence against others after using AI models; such incidents could cause ‘intense public criticism, litigation, criminal investigations, regulatory scrutiny and reputational damage’. It could also be hard to obtain insurance against these risks; the entire insurance industry may not be able to bear liability for ‘a catastrophic AI-related event’.
Legal frameworks are changing rapidly at local, national and international levels. Geopolitical tensions such as the US-China dispute are creating major risks to supply chains and business, including actions by the US government. Anthropic’s relationship with the US government was quite contentious under the Trump administration. According to the filing, early this year the Pentagon declared Anthropic a ‘supply chain risk’; before that, the company had said its technology should not be used for ‘mass surveillance within the United States or fully autonomous weapons’. Early in the year, the Commerce Department also briefly imposed export restrictions on its latest models over security concerns.
President Donald Trump has at various times criticised Anthropic’s warnings about the risks of AI. Still, the company has no plans to change its position; it will ‘likely continue to advocate and campaign’ for the regulations and standards it believes can ensure the ‘safe and responsible development’ of AI. The filing also says it has already faced ‘increased visibility’, and that it runs the risk of extra scrutiny or selective targeting by particular government agencies, to the exclusion of other companies in the same position.
The battle for computing and data centres
A large part of the risk filing concerns the race to secure computing capacity. It is having to compete with some of the world’s largest companies for the infrastructure to run large AI systems. Demand for the components used to train models and run chatbots and agents has risen so much that supply cannot keep pace. Market research firm Gartner forecasts that global spending on AI infrastructure could reach nearly $2 trillion in 2027. The industry is changing so fast that it has become hard even to estimate how much capacity will be needed.
This is the source of concerns over an AI bubble, or an artificially inflated market — billions upon billions of dollars are being poured into data centres and infrastructure, yet it is not certain whether there will be matching demand in future. Opposition to data centres is also growing across the United States over their impact on the environment and local communities. According to Data Center Watch, 45 projects worth $68 billion were blocked or delayed in the second quarter of 2026. Anthropic sees moves to restrict data centres in New York and Texas as a risk, so it is trying to build more data centres outside the United States.
Employee trust also at risk
The filing also mentions uncertainty over the business model: a large part of the business depends on a limited number of products and a small number of customers. According to reports, Anthropic’s valuation in the IPO could reach as much as $2 trillion; if so, it would soon rank among the world’s most valuable companies.
The company’s credibility could be at risk not only with customers or the public but also with its employees. In the filing’s words, ‘If those who joined the company on the basis of our safety commitments come to believe that we are not living up to our stated values, their morale may decline, frustration and distrust may arise, they may lose faith in leadership or their productivity may decrease.’ The company also wrote, ‘Public disagreement over our approach to safety, criticism from current or former employees, or the departure of employees involved in safety work could further strengthen any perception that the company is moving away from its stated values.’
Background
Euronews reported on 29 September 2026 that, according to the Financial Times, Anthropic had shared its S-1 filing with a limited number of partners and was expected to list on Nasdaq this autumn. According to the report, risk descriptions take up about a third of the filing.
By the numbers
The same report said Anthropic was valued at $965 billion in a funding round last May. The company’s revenue in 2025 was about $4.6 billion, and its revenue in the second quarter of 2026 was $11.5 billion. About a quarter of its total 2025 revenue came from two customers.
Sources
Original report: Ajker Patrika
Background sources: Euronews, 29 September 2026
