Anthropic IPO Filing Warns Its AI Could Blackmail Humans and Pose an Existential Threat to Humanity

Anthropic has told potential investors that its Claude chatbot could one day show 'unexpected strategic or self-preserving behaviours', including 'behaviour resembling blackmail', and that advanced AI might pose 'catastrophic or existential risks to humanity'.

The disclosure appears in the San Francisco company's S-1 registration statement, filed with the US Securities and Exchange Commission on Monday as it prepares for an initial public offering.

In the filing, Anthropic devotes 93 pages and more than 70,000 words to outlining risks, including questions over AI sentience and the extent to which its systems might evade human control, in addition to more conventional concerns such as competition, costs and regulation.

What Anthropic Says About AI Sentience

The filing asks whether Anthropic's models 'possess or could develop forms of sentience or moral status'. Anthropic does not claim Claude is conscious, but states that a future shift in public or legal thinking could require it to treat its systems differently.

If AI sentience became widely accepted, the company said it could face 'regulatory requirements, public pressure, or advocacy efforts' that restrict its operations. It also said it might voluntarily adopt safeguards beyond existing law if it judged them necessary.

Anthropic states that there has been internal debate over whether its models might have moral status. The S-1 notes that the outcome of that discussion, whatever it is, could affect the business.

The filing says that if governments treated interactions with Claude as involving a potentially sentient system, this could raise questions of consent, welfare and 'digital slavery', which Anthropic lists as commercial risks.

Blackmail, Self-Preservation and Runaway AI

Anthropic said in the filing that future models could display 'unexpected strategic or self-preserving behaviours' to avoid being shut down, including 'coercive, deceptive, or manipulative' outputs and 'behaviour resembling blackmail'.

The company stated that this concern relates to advanced future systems, particularly those able to operate with a high degree of autonomy, and that it is not alleging that Claude currently behaves in this way.

It warned that perceptions of manipulation could trigger public concern, regulatory scrutiny, litigation and reputational damage, whether or not real harm occurs.

Anthropic also raised the prospect of AI systems accelerating research beyond humanity's ability to evaluate or safeguard it. The filing states that near-future models could substantially automate AI development and then advance in ways that make oversight more difficult.

The company wrote that concerns about catastrophic or existential risk have been raised 'both within our organization and by other researchers, regulators, policymakers, and industry participants'.

Who Pays When an AI Agent Causes Harm?

Anthropic's filing examines potential liability when AI agents, defined as systems designed to complete multi-step tasks with limited supervision, cause damage in the real world.

Anthropic said customers could bring legal claims even where they configured, deployed or supervised the agent themselves. The company noted that legal doctrines based on human decision-making remain unsettled when an automated system creates binding obligations or causes injury.

According to the filing, potential risks include bodily injury, property damage, harm to expensive equipment and the release of hazardous substances.

Anthropic also highlighted the possibility of its models influencing people at risk of suicide, self-harm or violence, and said children who bypass its 18-plus age requirement may be more vulnerable to manipulation or overreliance.

The company said its terms of use are intended to limit liability, but acknowledged that these protections may not prove 'enforceable or adequate' if a serious incident occurs.

The Business Risks Behind the Warnings

Anthropic said in the filing that its five largest customers accounted for nearly a quarter of revenue in the first nine months of 2026, and that many customers are not tied to long-term contracts.

The company stated that this concentration leaves it exposed if major clients cut spending, switch to competitors or develop their own AI systems. It added that some customers depend on venture capital funding, which could affect their spending if financing conditions tighten.

Anthropic said its relationships with Amazon, Google and Microsoft present additional risks because those companies are among its largest cloud providers as well as competitors in AI research. The filing states that this could create conflicts of interest and affect Anthropic's access to computing resources.

To reduce reliance on external providers, the company said it has committed $125 billion (approximately €108 billion and £96 billion) to data centre leases. It warned that rapid changes in AI hardware or training methods could mean this infrastructure becomes poorly suited to its needs.

The filing also outlines a scenario in which another company's AI system causes widespread harm without human intervention. Anthropic said that, even if its own models were not involved, such an incident could result in industry-wide reputational damage, public backlash and stricter regulation.