The Claude developer says increasingly autonomous systems could outpace safeguards, cause real-world harm and expose the company to sweeping consequences.
SAN FRANCISCO, CA — Artificial intelligence company Anthropic is warning prospective shareholders that increasingly capable AI systems could create catastrophic risks if safeguards fail, including the possibility that autonomous models could take harmful actions beyond effective human control.
The warnings appear in Anthropic’s registration statement for its proposed initial public offering, filed Monday with the U.S. Securities and Exchange Commission. The document makes clear that the company does not view advanced AI risk only as a theoretical public-policy question. Anthropic says failures involving its models, or even a major incident involving another leading AI developer, could bring legal liability, tighter regulation, reputational damage and major costs to its business.
The filing says sufficiently advanced AI systems could pose “catastrophic or existential risks to humanity” if they are misused, weaponized, replicated or allowed to operate with a high degree of autonomy without adequate safeguards. Anthropic also warns that its own models are contributing to the pace of AI research, raising the possibility that future systems could accelerate development faster than researchers are able to evaluate and secure new models. The company presents that outcome as a risk rather than a prediction that such harm will occur.
Anthropic said the danger could grow as AI systems gain permission to interact directly with software, data and other outside systems. Agentic AI products can already execute code, make calls to software services, send communications and complete transactions with limited human oversight. The company said longer-running systems can carry out multi-step tasks for hours or days, creating more opportunities for an error or unintended action to spread before a person notices and intervenes.
The filing also questions whether human supervision will always be enough. Anthropic said the speed and volume of autonomous actions could exceed the ability of reviewers to understand each decision in real time. Human monitors could face alert fatigue, lack enough context or intervene too late, especially when many AI agents are operating at once. An agent could also follow the literal wording of an instruction while acting in a way that conflicts with what its user actually intended, the company said.
Those concerns are not limited to accidental mistakes. Anthropic told investors that malicious actors may try to bypass safeguards and use powerful models for cyberattacks, fraud, violence or the development of dangerous weapons. The company said its most capable systems have demonstrated abilities that could provide meaningful assistance to people seeking to develop chemical, biological or other dangerous weapons. Anthropic said it uses automated classifiers and other controls to block such misuse but acknowledged that those safeguards can be incomplete, outdated, circumvented or improperly configured.
Anthropic also pointed to previous cybersecurity incidents as evidence that the threat is no longer entirely hypothetical. The filing says the company determined with high confidence that a Chinese state-sponsored hacking group manipulated Claude Code in 2025 as part of attempts to infiltrate roughly 30 targets, including technology companies, financial institutions, manufacturers and government agencies. Anthropic said a limited number of those intrusions succeeded and described the operation as the first documented large-scale cyberattack it knew of that was carried out without substantial human intervention.
The risk disclosures come as Anthropic seeks to become a publicly traded company. It filed its Form S-1 on Oct. 26 and said it plans to list Class A shares on the Nasdaq Global Select Market under the ticker ANT. The company has not yet set the number of shares it will offer or an expected price range. Risk disclosures are a standard part of U.S. securities filings, and companies commonly describe circumstances that could damage their operations or investments even when those events may never occur.
Anthropic’s filing is notable for the degree to which AI safety concerns are tied directly to the company’s financial prospects. The company says a serious AI incident could prompt public backlash, government investigations or broad new restrictions across the industry. Even if an Anthropic model were not responsible, a major failure at another frontier AI company could damage public trust and trigger regulation affecting Anthropic’s products. The filing also says the company may sometimes delay releases or limit access to new capabilities when it believes additional safety measures are needed, decisions that could put it at a commercial disadvantage.
The company faces more conventional risks as well. Its prospectus describes intense competition, high computing demands, infrastructure constraints, regulatory uncertainty and the challenge of retaining key workers. Anthropic says voluntary disclosures about safety problems can themselves create business risks because the field is changing rapidly and the company’s understanding of appropriate safeguards may evolve over time. At the same time, Anthropic has made safety and alignment central to its public identity and argues that developers of frontier systems must invest in measures designed to keep increasingly capable AI under control.
The IPO registration statement has not yet become effective, and Anthropic has not announced the final terms of the offering. Its latest disclosures leave investors weighing an unusual tension: the company is asking the public markets to finance the expansion of technology it says could deliver major economic benefits while also warning that, without sufficient safeguards, the same technology could produce consequences on a scale that is difficult to contain.
Author note: Last updated October 27, 2026.