Anthropic IPO warns its AI could blackmail people, become sentient, end humanity

The SEC just published Anthropic’s S-1, and the filing did what these things always do: it invited everyone to stare at the money. Investors are watching the company’s fast-growing revenue and its even faster-growing costs. Anthropic, for its part, also tried to sell a future where AI “rapidly transform[s] every sector of the global economy” in a way bigger than industrialization, electricity, semiconductors, or the internet.

But the document spends a surprising amount of space on what could go wrong. The risk section runs 93 pages and more than 70,000 words, and it lays out hundreds of possible problems that could hit the business. Some are the usual corporate anxieties — competition, costs, customer acquisition, regulation. The rest are much stranger, and much more specific to a company trying to market frontier AI as both lucrative and safely manageable.

That tension is the story here. Anthropic wants public-market buyers to believe the company can ride the AI boom, but its own filing keeps circling back to the idea that the same technology could create its biggest problems. The S-1 doesn’t hide that. It practically underlines it.

Ars Technica’s read of the filing narrows the giant list to seven especially interesting risks, the kind that say as much about the current AI moment as they do about Anthropic. The result is a prospectus that feels less like a smooth pitch deck and more like a warning label attached to a gold rush.