Anthropic is asking investors to back its stock market debut. In the same document, it admits that if its AI ever caused a genuine catastrophe, the insurance industry might not be able to handle the bill.
The San Francisco company behind the Claude models filed its Form S-1 with the Securities and Exchange Commission on Monday. It plans to list on the Nasdaq Global Select Market under the ticker ANT. Anthropic has confirmed the filing, though it hasn’t yet settled on a share count or price range.
Underwriters will want to read one passage in the 94-page risk factors section really closely. Anthropic says the damage from a catastrophic AI-related incident could exceed what any single insurer, or possibly the market as a whole, has the capacity or the expertise to underwrite. It describes the market for AI-specific cover as nascent, and says it may not be able to buy protection for some AI liabilities on commercially reasonable terms, if it can buy it at all.
Read next: AI exclusions are spreading through commercial policies, brokers warn clients to check cover
Much of the filing deals with the kind of risk carriers find hardest to price. As AI systems take on more work with less human oversight, Anthropic says, they could end up taking harmful actions independently and at scale, with nobody stepping in. It puts that down to a few possible causes: misalignment with what people intended, mistakes in how objectives are set, or self-preserving behavior.
How widely Claude is used makes this more pressing. As of September 30, the companies paying Anthropic at least $100,000 a year included every member of the Fortune 10, more than half of the Fortune 500 and over 30% of the Global 2000. If a model that popular failed, or was turned against the businesses using it, a large part of a commercial book could feel the effects at the same time.
There’s a supply chain issue as well. Anthropic depends on Amazon, Google and Microsoft for much of the computing power behind its models, and it says its own insurance may not cover the business interruption losses that a serious outage would cause. Clients that have built Claude into their day-to-day operations might reasonably ask whether their contingent BI cover would respond in that situation. Many current policies may not give them a clear answer.
“I do strongly believe that there's going to be an interesting market for, for AI liability,” Daniel Elechiguerra Batlle told Insurance Business. “Especially for the autonomous use cases that are completely being almost ignored right now by insurers.
On cyber, the filing points out that criminals are already using AI to find vulnerabilities, write malware and get past detection tools. It also concedes that Anthropic’s own cyber cover may fall short, with claims potentially going unpaid or being paid late.
Read next: AI exclusions are creeping into insurance – but cyber policies aren’t the issue yet
Some of Anthropic’s exposure is already in its contracts. Several agreements with the outside firms that test its models require Anthropic to defend and indemnify them against certain third-party claims with no cap on its liability. Some customer contracts include similar uncapped promises for intellectual property claims over model outputs.
The company has also already seen what happens when cover runs short. In September 2025 it agreed to pay $1.5bn to settle the Bartz copyright class action, a deal a federal court approved in July. Anthropic says its insurance didn’t fully cover it.
The regulatory picture is getting busier too. In October the FTC sent Anthropic a civil investigative demand asking about its consumer protection, data protection and privacy practices, and about the safety and security of its products. That followed September media reports, citing a senior FTC official, that the agency had opened non-public investigations into the risks increasingly autonomous AI poses to consumers.
States are moving as well. The prospectus names New York’s Responsible AI Safety and Education Act and California’s Transparency in Frontier Artificial Intelligence Act, both of which require frontier model developers to report incidents. Anthropic publicly endorsed the California law, but it warns that a late, incomplete or inaccurate disclosure could still bring compliance and liability exposure. Meanwhile, it is fighting the federal government in court over February’s moves to stop agencies using its technology and to label it a supply-chain risk. Those moves came after it refused to let its models be used for mass domestic surveillance or fully autonomous weapons.
Anthropic expects the costs of being a public company to push up its D&O premiums. It says it may have to settle for lower limits or pay substantially more for similar cover.
Its structure is also unusual. Anthropic is a public benefit corporation, so its board has to weigh stockholder returns against a stated public mission. An independent Long-Term Benefit Trust, which has no financial stake in the business, can elect a majority of the directors, which is why Nasdaq treats the company as a controlled company. On top of that, the charter says that failing to strike that public benefit balance won’t, by itself, count as bad faith or a breach of the duty of loyalty when it comes to indemnifying directors and officers. Most D&O wordings weren’t written with a clause like that in mind.
Since January, US carriers have been able to bolt three optional ISO generative AI exclusion endorsements onto commercial general liability policies. Some London specialists, CFC among them, have gone the other way and written affirmative AI language into their products. For a US broker placing business in both markets, the same AI-related loss could be excluded on a domestic CGL form and covered under a London policy.
Insurers are also customers, of course. The prospectus lists insurance underwriting among the production work customers run on Claude, and Tampa-based broker the Baldwin Group is among those that have widened their use of the platform this year.
Read next: Baldwin Group expands AI partnership with Anthropic
