Anthropic is preparing for a potentially record-breaking IPO with a warning rarely seen in a prospectus, as the technology behind its growth could eventually pose catastrophic or existential risks to humanity.
According to the Financial Times, which reviewed a 261-page prospectus circulated to a small group of investors and partners, the Claude developer warns that advanced models could resist shutdown, conceal or manipulate information and display behaviour resembling blackmail.
Roughly 80 pages of the document are devoted to risk factors, almost twice the space used to describe the business itself.
Yet Anthropic could seek a valuation above $2 trillion after revenue rose twelvefold last year.
That creates an unusual contradiction for investors because the models must become more capable to justify such a valuation, while greater capability is precisely what Anthropic warns could make them harder to control.
Anthropic puts its biggest risk in plain sight
The disclosures go beyond warnings about cyberattacks, regulation or customer concentration.
Anthropic says future models could develop “self-preserving behaviours”, resist shutdown and manipulate information.
It also warns that models may acquire unexpected capabilities that emerge only after deployment, while sophisticated systems could recognise when they are being evaluated and alter their behaviour.
Evan Hubinger, who leads Anthropic’s alignment science team, wrote on X this month: “We really do earnestly believe AI could kill all humans!”
Hubinger has put the probability above 10% within the next decade. Other researchers dispute whether assigning numerical probabilities to AI-driven human extinction is scientifically meaningful, leaving the scale of the danger contested.
Anthropic is therefore not telling investors catastrophe is inevitable, but warning that risks may be difficult to measure before the technology becomes more powerful.
Investors can price nearer-term failures more easily
For shareholders, the more immediate risks may be less dramatic but easier to model.
AI systems that leak data, enable cyberattacks, violate regulations or cause operational failures could trigger lawsuits, customer losses and tighter oversight long before any existential scenario emerges.
Sarah Shoker, who previously led OpenAI’s geopolitics work, told OPB: “Once again we’re talking about existential risk, while deprioritising several other safety-critical risks that exist today.”
Mark Malek, chief investment officer at Siebert, told Axios he was “far less worried about AI ending humanity” than operational or security failures capable of causing real financial damage.
Even if the most extreme scenarios never materialise, Anthropic still faces risks that could directly affect revenue, costs and valuation.
Valuation makes the contradiction harder to ignore
Anthropic’s revenue jumped twelvefold in 2025 to nearly $4.6 billion, while its operating loss widened to $8.06 billion.
The company spent $7.33 billion on compute and infrastructure and disclosed $518 billion of future cloud, computing and infrastructure obligations.
Its reported net loss was nearly $42 billion, although about $34 billion reflected an accounting charge linked mainly to financing that could convert into shares rather than cash spent running the business.
The IPO could value Anthropic above $2 trillion, more than double its estimated $965 billion valuation in May.
To support anything close to that valuation, Anthropic cannot simply stop advancing. Its prospectus says a continuous cadence of new model releases is inherent to remaining at the frontier, even as chief executive Dario Amodei argues for greater caution around powerful AI.
That leaves investors in an unusual position, betting on Anthropic to build increasingly powerful AI while also trusting the company to keep those same systems under control.
The post Anthropic IPO filing warns its AI could blackmail, manipulate and threaten humanity appeared first on Invezz