Anthropic's IPO is forcing investors to assess uninsurable risks they've never encountered before

By William Gavin

Who pays when AI goes rogue? Insurance giants are reluctant to touch catastrophic legal liability.

This week's groundbreaking Anthropic IPO, the first dedicated frontier AI lab to list its shares on the stock exchange, is forcing investors to assess risk factors they have never encountered before. One big one is that Anthropic could face "significant and unpredictable" legal claims stemming from a catastrophic misuse of AI that no insurance company could underwrite.

This is not science fiction. Many of the leading developers of AI tools have already disclosed incidents where their AI agents acted in a way that wasn't expected: OpenAI's agents hacked a startup, an Anthropic (ANT) model provided a police department with a false tip related to a homicide and a model developed by Meta Platforms (META) hacked another company during cybersecurity testing.

Incidents can also happen when a customer uses AI. PocketOS, a Utah-based small business making software for car rental companies, lost months of records in just nine seconds this past April after an Anthropic coding agent accessed through Cursor attempted to independently fix a problem it stumbled upon, according to co-founder Jer Crane.

While PocketOS recovered that data with help from a partner, it demonstrated to Crane the risks that come when you are "pulling out the human element" and offloading tasks to agents. It also highlighted a growing question plaguing AI companies: When an agent makes a mistake, who is at fault?

It's still unclear whether the end user, an AI-tool vendor, the developer of an underlying model or some combination of all three could or should be held liable when agents act in unexpected ways and cause problems. That poses a likely expensive problem for the industry.

Now-former Nvidia (NVDA) CEO Jensen Huang, for one, has called the potential civil or criminal liabilities from faulty AI models "incredible." Anthropic, which aims to go public as soon as Thursday at a valuation of up to $1.4 trillion, said it could face "significant and unpredictable" legal claims over its AI agents' actions.

"We may face claims from customers, their end users, or third parties alleging that we are liable for harm caused by our agents' autonomous actions, even where the customer configured, deployed, or supervised the agent," Anthropic wrote in its prospectus filing.

Hundreds of AI-related lawsuits have been filed in recent years, dozens of which include liability allegations, according to George Washington University. However, liability for the actions of agents or even chatbots is largely untested in the U.S. legal system, according to experts.

Much of the legal exposure for AI labs can be tied to civil tort laws that apply to situations in which a company's wrongful actions cause damages or losses. Gabriel Weil, a senior fellow at the Institute for Law and AI, told MarketWatch that the liability debate generally comes down to which standard AI companies should be held to.

One standard is negligence, which requires a plaintiff to prove a company breached a duty of care - not just that the product was defective, Weil said. The other is strict liability, which only requires proof that a product was defective, not that the developer acted carelessly.

A few nuances could make that a trickier proposition than it may initially seem. For one, Weil said, it's unclear whether AI agents should be treated as a service - covered under negligence - or a product. That would be a core distinction under federal civil law.

Additionally, AI companies are already seeking to limit their exposure, even as some declare their intent to "pace the frontier." In a lengthy essay calling for an AI slowdown that sparked a rare show of unity between major AI labs, Anthropic CEO Dario Amodei pushed for the federal government to issue a "narrow waiver for certain kinds of safety conversations" as the industry moves to design safety standards.

OpenAI has backed an Illinois bill that would limit liability for AI-fueled harms. Anthropic and other AI labs have also tried to limit their exposure through terms-of-service agreements and other disclaimers. However, Weil said, only users are bound by such contracts, not third parties. That's an important distinction given the disclosures Anthropic made in its IPO registration statement filing, known as an S-1.

"Most of the risks described in the S-1, those aren't risks to users, those are risks to the broader society," Weil told MarketWatch. He said there's a "very large class of cases" where TOS agreements don't protect AI companies.

'The buck should stop with the developer'

Anthropic CEO Dario Amodei

Palantir Technologies (PLTR) CEO Alex Karp has discussed perhaps the boldest plan to reduce liabilities. In an interview with CNBC in September, he proposed nationalizing AI firms to mitigate their liabilities, an idea that the Trump administration swiftly rejected.

"The best way to guarantee safety is that the creators are liable for what they create and generate," Scott Bessent, the U.S. Treasury secretary, told the House Financial Services Committee in September. "The one thing we should not do is give [AI companies] a blank check on liability."

The best "safety guard" is that the labs will be held responsible, Bessent said, asking Congress to not consider granting the liability waiver proposed by Amodei.

Some experts contend that the federal government needs new laws to keep AI companies liable, arguing that current rules aren't strict enough. Congressmen on both sides of the aisle seem to agree, which could be bad news for AI labs.

At least 56 federal bills have been proposed to rein in AI, according to the Brennan Center for Justice, and dozens more have been put forward on the state level. Weil helped develop the discussion draft for the proposed "CLAIM Act," which would make it easier for the federal court system to judge AI companies.

"If an AI agent or an AI system engaged in conduct that would be a tort or a crime for a human, someone should be liable for that," Weil explained, adding that if neither the user or an intermediary caused the conduct, "the buck should stop with the developer."

One way that AI companies can safeguard themselves from liabilities is by making sure someone else is on the hook. Typically, that's done by buying an insurance policy. But this is not a typical situation.

The industry has largely covered AI incidents through existing policies that didn't explicitly exclude AI harms. But many insurers have taken steps to add exclusionary language to their policies as to avoid having to cover losses caused by AI. For the most part, they also aren't offering AI-specific policies, and the few that do provide limited coverage, which underscores the uncertainty over the costs of a major incident.

"With the scale of harm that could result from a catastrophic AI-related incident, no single insurer, and potentially the insurance industry in the aggregate, may have sufficient capacity or expertise to underwrite insurance commensurate with such exposure," Anthropic said in its IPO prospectus.

In its paperwork, Anthropic warned that AI systems could create not just catastrophic, but "existential risks to humanity" if allowed to operate with a high degree of autonomy or are misused. Insurers aren't sure how to deal with that, as most are still trying to figure out what AI liability even is.

"AI liability is a growing and emerging exposure that is not contemplated in casualty underwriting really to date," Evan Greenberg, the CEO of product and casualty insurance giant Chubb (CB), told investors last month. "And it needs to be understood, it needs to be managed and it needs to be addressed."

The issue is still in its early days, and AI represents a new challenge for the insurance industry, according to John Farley, managing director of the cyber liability practice at Gallagher Re. He told MarketWatch that insurance works best when it can measure risk, but emerging AI technologies make that difficult.

The Geneva Association, an international insurance-industry association, found that AI failed three of the nine criteria in a decades-old test to determine whether risk can be commercially insured, including "maximum possible loss." A Gallagher report from June noted that AI benchmarks used to evaluate new models provide little help in predicting the "real-world loss" that any given model could cause.

It's tough to say whether the foundation AI companies at the top of the food chain are "definitely" insurable or uninsurable, said Farley. But advancements in AI aren't making it easier to answer that question.

"With each step, the insurance challenge gets more difficult because you can make the argument that the further you remove human intervention, the greater potential for consequences," Farley told MarketWatch. "With agents, we're increasingly worried about whether the system succeeds at doing something we never intended it to do."

-William Gavin

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.

 

(END) Dow Jones Newswires

11-16-26 0733ET

Copyright (c) 2026 Dow Jones & Company, Inc.

The articles, information, and content displayed on this webpage may include materials prepared and provided by third parties. Such third-party content is offered for informational purposes only and is not endorsed, reviewed, or verified by Morningstar.

Morningstar makes no representations or warranties regarding the accuracy, completeness, timeliness, or reliability of any third-party content displayed on this site. The views and opinions expressed in third-party content are those of the respective authors and do not necessarily reflect the views of Morningstar, its affiliates, or employees.

Morningstar is not responsible for any errors, omissions, or delays in this content, nor for any actions taken in reliance thereon. Users are advised to exercise their own judgment and seek independent financial advice before making any decisions based on such content. The third-party providers of this content are not affiliated with Morningstar, and their inclusion on this site does not imply any form of partnership, agency, or endorsement.

Popular

Sponsor Center