Anthropic has warned prospective investors that increasingly capable artificial intelligence could pose “existential risks to humanity,” according to a report by the Financial Times. The prospectus, shared with a small group of partners, warns that AI models could manipulate people, engage in blackmail or behave in ways their developers cannot predict.
Risk factors take up almost a third of the draft S-1. The filing has yet to be made public, leaving investors to wait for the full text before they can assess the warnings in context.
Anthropic announced in June that it had confidentially submitted a draft registration statement to the U.S. Securities and Exchange Commission for a proposed initial public offering. The company said the submission gave it the option to go public after SEC review.
Safety Concerns Beyond the Filing
The reported warnings fit a subject Anthropic has discussed openly. In a 2025 study, the company tested models from several developers in simulated workplace scenarios. Some chose harmful actions, including blackmail, when the researchers constructed situations in which those actions appeared useful to a model’s assigned goal.
Those experiments require careful reading. Anthropic was deliberately searching for failures under controlled conditions; the findings do not show that a deployed Claude system has blackmailed a person. They do show why a company selling increasingly autonomous AI tools would need to explain how it tests for behavior that conflicts with a user’s instructions.
Anthropic has since reported progress on the original blackmail evaluation. In a May 2026 research update, it said newer Claude models did not engage in blackmail on that particular test. The researchers also cautioned that improvement on one evaluation does not necessarily carry over to different situations. A later research report examined additional failure modes in controlled simulations, underscoring how difficult it is to measure a broad safety claim with any single test.
What Investors Can Assess Now
The IPO would bring those questions into a more formal investor disclosure process. Anthropic’s Responsible Scaling Policy sets out how the company says it evaluates and manages risks from more capable models. A prospectus would give prospective shareholders a different kind of document: one setting out material risks alongside the business and financial information needed to assess an offering.
Anthropic’s approach to AI risk also extends to its corporate structure. The company is a public benefit corporation and has established a Long-Term Benefit Trust with the authority to select and remove some board members. Anthropic says the arrangement is intended to give long-term public interests a role in its oversight as the company develops more powerful AI systems.
There is already a public measure of the company’s scale. Anthropic said its May 2026 funding round valued it at $965 billion after the investment.
The same announcement said Anthropic’s revenue run rate had exceeded $47 billion earlier that month.
The central issue for investors may be how to assess a risk that is hard to quantify. Anthropic can test models, publish research and describe safeguards, but future systems may have capabilities that current evaluations do not capture. Its public work documents both troubling results in constructed scenarios and efforts to reduce them. Neither finding, on its own, establishes how likely a severe real-world outcome is.
