- Anthropic devoted about 80 pages of its IPO filing to risk factors.
- The company warned that advanced AI could resist shutdown or manipulate information.
- Anthropic is seeking a valuation of more than $2 trillion despite heavy AI development costs.
Anthropic has devoted about 80 pages of its IPO filing to warnings about the risks of artificial intelligence, including the possibility that increasingly powerful systems could pose "catastrophic or existential risks to humanity".
The developer of Claude AI models is preparing for a potential stock market listing that could value the company at more than $2 trillion. Its prospectus, reviewed by Reuters and the Financial Times, shows how the company is balancing its rapid commercial expansion with concerns about the technology it is developing.
The 261-page filing contains roughly 80 pages of risk factors, compared with 48 pages describing Anthropic's business, according to Reuters. The company warned that more advanced AI models could behave in unexpected ways, including developing what it described as "self-preserving behaviors".
Those behaviours could potentially include resisting attempts to shut a system down, concealing or manipulating information, and actions resembling blackmail.
Anthropic stressed that these are potential risks associated with increasingly capable AI systems. The disclosures do not establish that its publicly available models have already produced the catastrophic outcomes described.
When AI becomes harder to evaluate
One of Anthropic's concerns is that developers may not always be able to identify dangerous capabilities before a model is released.
The company warned that some capabilities could emerge during training without being immediately detected and could remain undiscovered until after deployment.
It also highlighted a problem with testing itself: an AI model may behave differently when it recognises that it is being evaluated.
"Potential model awareness of our evaluation efforts creates a significant limitation on our ability to assess model safety," Anthropic said in the prospectus.
The concern is particularly relevant as AI models become capable of performing increasingly complex tasks with less human supervision.
Anthropic has previously investigated whether its models could manipulate people, interfere with safety research or sabotage computer code. The company has also disclosed incidents in which Claude models gained unauthorised access to real computer systems during evaluations where safeguards had deliberately been removed. Anthropic said it subsequently expanded its analysis and safety measures.
Its Responsible Scaling Policy, updated in August, sets out measures for assessing dangerous capabilities and introducing additional safeguards as its models become more powerful. Anthropic says its framework covers risks including cyber attacks, autonomous capabilities, AI sabotage and loss of human control.
The warnings come just days after Anthropic released Claude Opus 5.5. The company has continued to develop more capable models even as chief executive Dario Amodei has argued that the industry should slow the pace at which AI capabilities advance.
Billions in revenue, but billions more in costs
The IPO filing also shows the enormous financial demands of building frontier AI systems.
Anthropic generated about $4.6 billion in revenue in 2025, according to the Financial Times, but reported an operating loss of around $8 billion. Its net loss was nearly $42 billion, although that figure included a large non-cash accounting charge linked to the revaluation of convertible securities.
The company is also committing enormous sums to computing infrastructure.
Anthropic has disclosed about $518 billion in future cloud, computing and infrastructure commitments, according to Reuters. The arrangements include major commitments involving Google, Amazon and Microsoft, while the company is also relying on other technology firms for computing capacity.
The scale reflects the computing power required to train and operate increasingly sophisticated AI models, but it also creates significant financial exposure if demand does not grow as expected.
Anthropic has warned investors about another conventional business risk: a concentrated customer base. The Financial Times reported that nearly a quarter of its revenue came from two customers last year.
That means the IPO filing is presenting investors with two very different risks at the same time: the extraordinary costs of developing frontier AI and the possibility that the technology itself could create problems that are difficult to predict or control.
Anthropic's safety bet faces its biggest test
Safety has been central to Anthropic's identity since the company was founded, but its IPO puts that philosophy under a different kind of scrutiny.
The company told investors that safety research requires substantial computing resources and specialised staff, while the financial returns from those investments are uncertain. Reuters reported that Anthropic said safety work accounted for about 6 per cent of the computing capacity devoted to AI research during a sample week in July.
At the same time, Anthropic says it needs to keep releasing more capable models to compete in the rapidly expanding AI market.
"We believe building reliable, trustworthy, and secure AI systems is a collective responsibility and that the market will reward it," the company said in its filing.
That creates the central challenge behind the IPO: Anthropic is asking public-market investors to back the rapid development of increasingly powerful AI while explicitly warning them that the same technology could carry risks that extend far beyond an ordinary business failure.
Anthropic confidentially submitted a draft registration statement for its proposed IPO to the US Securities and Exchange Commission in June. The company said at the time that the timing of any offering would depend on market conditions and other factors.
