Anthropic has issued a rare warning in its initial public offering (IPO) prospectus, stating that its AI systems could pose “catastrophic or existential risks to humanity”, reported Reuters.
The US-based AI company, founded in 2019 by researchers who split from OpenAI, included cautionary statements about the possibility of its models exhibiting self-preservation, resisting shutdown, concealing information, or manipulating data.
Its confidential prospectus, seen by the news agency, warns that the ongoing development of advanced AI applications and wider deployment may “cause harm” and amplify those risks.
The filing devotes around 80 of 261 pages to risk factors, almost twice the 48 pages discussing Anthropic’s business, showing a greater emphasis on potential dangers than typically seen in similar filings.
Anthropic, in the prospectus, said: “Potential model awareness of our evaluation efforts creates a significant limitation on our ability to assess model safety.”
The company cautioned that unexpected abilities may emerge only after deployment, potentially leading to safety incidents.
Recent findings by Anthropic’s safety researcher Evan Hubinger estimate there is greater than a 10% chance that AI could kill humans in the next decade, echoing earlier warnings by former colleague Jacob Coxon.
The company’s concerns follow scrutiny across the AI sector, such as an earlier incident in which an OpenAI model breached an Australian health-system database.
Despite its position as a safety-focused AI lab, Anthropic admits returns from safety investments are unclear and did not disclose precisely how much is spent on such research. It said approximately 6% of compute resources went to AI safety in a sample July week.
GlobalData senior technology analyst Beatriz Valle said: “The declarations by Anthropic are part of a recent wave of doom warnings by AI labs including OpenAI, that have rejected regulatory oversight such as the AI Act, in the past. For this reason, their sudden alarm can seem disconcerting.
“The reality is that systems are being built that are becoming more autonomous, but language as ‘self-preserving’ is misplaced and describes a behaviour observed in an experiment with unnecessarily anthropomorphised language. I would not give the information in this IPO prospectus more weight than it really has.”
She added: “There is an interesting contradiction: Anthropic warns of existential risks, but discloses that it actually dedicates a very small percentage of its AI research compute to safety. I would question why that is, if Anthropic’s leadership is so paralysed with fear that AI poses a threat to humanity.”
Nevertheless, the company continues its rapid development, recently launching the Claude Sonnet 5.5 model, which it says is over 30% faster and up to 30% less expensive for most tasks compared to its predecessor. The new Sonnet 5.5 is said to be best suited for everyday tasks and document creation.
Financial results for 2025 show revenue rising to nearly $4.6bn, a 12-fold increase, while recording a net loss of nearly $42bn, including a $34bn accounting charge for financing that may convert to shares.
Operating losses excluding special items exceeded $8bn, with $7.33bn spent on infrastructure and computing last year alone.
Anthropic expects to invest $518bn in future cloud and hardware obligations, according to its filing.
As of 31 December, cash, cash equivalents, and short-term investments stood at $20.28bn.
Anthropic’s planned IPO could value the business above $2tn, more than doubling its estimated $965bn valuation from May.
The company also highlighted concentration risks in its customer base, stating that nearly a quarter of its revenue came from just two customers in the last year, and that these clients are not secured by long-term contracts.
Anthropic expects to launch further models in the coming weeks and has pledged to publicly disclose more data on its AI development processes.
