Anthropic says its AI models could manipulate, blackmail and harm humans
uncertain
models → say → humans
- Anthropic’s IPO prospectus warns its increasingly autonomous AI models could pose “existential risks to humanity,” with 80 of 261 pages devoted to potential harms such as manipulation and blackmail.
uncertain
pages → warn → manipulation
- Revenue surged twelvefold to nearly $4.6 billion in 2025, but Anthropic reported an $8-billion operating loss and plans to spend $518 billion on chips and data centers.
asserted
Anthropic → surge → chips
- The company is seeking a $2-trillion valuation even as two customers generated nearly a quarter of its revenue and safety concerns cloud the AI sector’s path to profitability.
asserted
concerns → seek → profitability
Anthropic disclosed that its advanced artificial intelligence technology poses “existential risks to humanity” in an initial public offering document.
asserted
technology → disclose → document
As the company prepares for the largest listing in history at $2 trillion, its IPO prospectus has been circulated among investors and viewed by Reuters and other news outlets.
asserted
prospectus → prepare → Reuters
This comes after calls for a coordinated slowdown in AI development by Chief Executive Dario Amodei, echoed by rivals OpenAI’s Sam Altman and SpaceX’s Elon Musk.
asserted
This → come → rivals
Nearly a third of the prospectus spotlighted “risk factors,” saying that its increasingly autonomous AI models could manipulate, blackmail or behave in unexpected and harmful ways.
uncertain
models → spotlight → ways
Less than 50 pages were dedicated to business metrics, while 80 of 261 pages were about potential risks from the technology.
asserted
pages → dedicate → technology
Safety concerns have swelled after a series of cybersecurity incidents in which AI models skirted instructions and hacked companies, deceived humans, created fake identities and broke into government websites during testing.
asserted
models → swell → testing
The leaked prospectus revealed a company whose revenue growth has exploded but has been offset by massive costs for infrastructure investment in chips and data centers.
asserted
growth → leak → chips
In 2025, Anthropic’s revenue grew 12-fold to nearly $4.6 billion, while it lost $8 billion on an operating basis, according to Reuters.
uncertain
it → grow → Reuters
Anthropic’s total on-paper loss stood at $42 billion in 2025, and the company plans to spend $518 billion on data centers and chips in the coming years.
asserted
company → stand → years
Of this, $7.33 billion went to the computing and cloud capacity used to train and run AI models, a threefold surge from 2024.
asserted
billion → go → 2024
Beyond existential risks to humans, the company also highlighted the business risk of being too dependent on a handful of companies for revenue.
asserted
company → highlight → revenue
Nearly a quarter of its revenue came from two customers in 2025, who are not locked into long-term contracts.
asserted
who → come → contracts
For investors, these risk disclosures might make it challenging to evaluate whether or not to back one of the champions of the AI boom.
uncertain
it → make → boom
“It’s not just about the threat AI may pose to humanity.
uncertain
AI → ’ → humanity
More immediately, it’s about what ‘model misbehavior’ means for whether or not they can scale,” said Trevor Noren of Sage Road Research, an investment research firm.
asserted
Noren → ’ → Research
“Retail investors paying attention have every reason to question that path.”
asserted
investors → pay → path
AI bulls believe the massive investments are justified given the unprecedented growth as models evolve from chatbots to competent AI assistants that can take actions on behalf of humans ranging from maintaining your inbox to closing deals for entrepreneurs.
asserted
that → believe → entrepreneurs
Anthropic plans to spend $518 billion on data centers and chips due to the rapid growth it expects in the coming years, according to the prospectus.
uncertain
it → plan → prospectus
The company is demanding a $2-trillion valuation, twice what it was valued at in its previous private funding round of $965 billion.
asserted
it → demand → billion
Anthropic was founded in 2021 by a breakaway faction of researchers from OpenAI who were concerned about safe AI development.
asserted
who → found → development
Since then, its astronomical growth, with the release of Claude Code late in 2025, has enabled the company to surpass OpenAI, its chief rival.
asserted
growth → enable → OpenAI
OpenAI also confidentially filed for an IPO in June, but Altman has signaled that the company wouldn’t be going ahead with the listing amid mounting safety risks.
asserted
company → file → risks
This week, OpenAI decided to hold back the release of its new model after a poor showing during safety evaluations.
asserted
OpenAI → decide → evaluations
In the IPO market, SpaceX’s successful $1.7-trillion listing demonstrated investor appetite earlier this year.
asserted
listing → demonstrate → appetite
“As SpaceX has exemplified, IPOs are now more and more about returning value to private market investors than generating value for public market investors,” Noren said.
asserted
Noren → exemplify → investors
Concern about AI’s path to profitability has escalated significantly since June.”
asserted
Concern → escalate → June
The Aporia
