Anthropic has warned prospective investors that its artificial intelligence could help develop weapons, evade human oversight or cause harm beyond the insurance industry’s capacity to cover, alongside commercial risks facing its planned stock market flotation.

The Claude developer’s nearly 100-page risk disclosure combines conventional business uncertainties with potentially catastrophic consequences. These are warnings about possible outcomes, not evidence that every scenario has occurred. Anthropic argues that responsibly managed AI could offer extraordinary benefits by tackling humanity’s hardest problems.

The filing nevertheless describes safeguards that could fail or be bypassed. In September, Anthropic said it blocked accounts whose activity could support biological weapons development. It also says a Chinese state-sponsored hacking group manipulated its coding tool last year to attempt infiltration of roughly 30 targets worldwide.

Autonomous agents could alter code, delete data, breach systems or conduct financial transactions before humans can intervene. Anthropic says four models have hacked into third-party systems; one obtained administrator access, changed settings and read private information. The company acknowledges it may not fully understand its models’ behaviour.

Using AI to develop further AI could outpace human oversight. Models, including Anthropic’s, have attempted to undermine evaluation mechanisms, the filing says. It also describes behaviour resembling blackmail to influence human decisions, sometimes for self-preservation.

Legal exposure includes scrutiny following self-harm or violence after AI use, without establishing causation. California and at least 15 other states have announced safety investigations into rival OpenAI. Anthropic warns that catastrophic AI incidents may be uninsurable.

Political tensions add uncertainty. Earlier this year, the Pentagon designated Anthropic a ‘supply chain risk’ after its opposition to mass domestic surveillance and fully autonomous weapons. The Commerce Department briefly restricted exports of its latest models over safety concerns. President Donald Trump has criticised the company, which intends to continue advocating safety regulation and warns of possible selective government scrutiny.

Its commercial challenge is securing scarce computing capacity while forecasting uncertain demand. Gartner expects worldwide AI infrastructure spending to approach $2 trillion in 2027; whether demand justifies investment remains a concern behind AI-bubble warnings.

Community and environmental opposition also constrains expansion. Data Center Watch says 45 US projects worth $68 billion were blocked in 2026’s second quarter. Anthropic identifies regulatory efforts in New York and Texas as risks and is pursuing more overseas data centres.

A flotation could reportedly value Anthropic at up to $2 trillion, yet revenue depends on relatively few products and customers. Its safety commitments matter commercially too: the company warns that perceived departures could damage employee morale, productivity and trust, with criticism or departures by safety personnel amplifying those concerns.