Anthropic has issued a stark warning. Its advanced AI models could pose grave risks. The warning was made in a draft IPO prospectus. The document was reviewed by Reuters.
Anthropic said advanced AI could create “catastrophic or existential risks to humanity”. The company also warned that some models may show unexpected behaviour during testing.
Models Could Resist Shutdown
Concerns were raised about model behaviour. Some systems were found to act in unexpected ways.
Anthropic said models could show “self-preserving behaviours”. They could also attempt to resist shutdown, conceal information or manipulate it, according to the prospectus.
Behaviour resembling blackmail was also cited. The examples were linked to controlled tests.
Anthropic said advanced systems may develop capabilities that are not seen during training. Such behaviour may only become clear after deployment.
Safety Takes Centre Stage
The warning forms a large part of the filing. About 80 pages focus on risk factors. The prospectus has 261 main pages. The business section covers about 48 pages.
Anthropic has presented itself as a safety-focused AI company. Yet it acknowledged that safety work is costly and difficult.
The company said safety efforts must compete for resources. Money is also needed for computing power and AI talent.
About 6 per cent of its computing power was used for safety work during one sample week in July, according to Anthropic.
A Huge Financial Bet
The safety warning comes as Anthropic seeks public investors. The company is preparing for a possible stock market debut.
Its revenue rose sharply in 2025. It reached nearly $4.6 billion. But heavy costs were also recorded. An operating loss of more than $8 billion was reported.
A net loss of about $42 billion was also recorded. Much of that figure came from a roughly $34 billion accounting charge linked to financing instruments.
The company has also made huge infrastructure commitments.
About $518 billion is planned for future cloud and computing needs.
Anthropic spent $7.33 billion on computing and infrastructure in 2025. That was more than half of its total operating expenses.
IPO Could Top $2 Trillion
The planned listing could draw major investor attention. Anthropic could be valued above $2 trillion. That would mark a sharp rise. The company was valued at about $965 billion in May.
The IPO would also put public investors directly into the AI race. Rival OpenAI is preparing its own possible listing.
Anthropic said reliable AI systems will matter to investors. The company described AI safety as a shared responsibility.
“We believe building reliable, trustworthy, and secure AI systems is a collective responsibility,” the company said in the prospectus.
Debate Over AI Safety
The warnings come amid wider safety concerns. Other AI companies have faced similar scrutiny.
Anthropic chief executive Dario Amodei has called for caution. He has argued that the pace of AI development should be slowed.
At the same time, new models continue to be released. The competition among leading AI firms remains intense. The central question is now becoming harder to avoid. How fast should increasingly capable AI be developed?
Anthropic’s filing offers no simple answer. But it puts the risks before potential investors in unusually direct terms.
About the Author
