Anthropic Says No Insurer May Have Capacity for a Catastrophic AI Loss

Anthropic has told prospective shareholders that a catastrophic incident involving its AI models could produce losses that "no single insurer, and potentially the insurance industry in the aggregate" has the capacity or expertise to underwrite, in a registration statement filed with the Securities and Exchange Commission on October 26.

The warning sits in the risk factors of the Form S-1 for the company's planned listing on the Nasdaq Global Select Market under the ticker ANT. No price range or share count has been set. Anthropic makes the Claude models; its tools are among those used in drafting at actuary.info.

The filing arrives as carriers are still deciding where AI exposure belongs. OpenAI holds up to $300 million of AI-risk cover placed through Aon, according to a Center for Strategic and International Studies report earlier this year, and both labs have looked at captives as general liability and cyber forms add AI exclusions. Anthropic's S-1 is the first time a frontier developer has put its own view of insurability into a securities filing.

Key Takeaways

Anthropic's S-1 says its policies "may not provide sufficient coverage" where one incident or a group of incidents causes disproportionate harm, and that a catastrophic AI incident may exceed the capacity of any single insurer or the industry together.
The $1.5 billion Bartz copyright settlement, expensed in 2025, was not fully covered by insurance, the filing says, and the company expects claims of that kind to recur.
Some agreements with external model evaluators, and some customer contracts covering model output, carry indemnities with no cap on Anthropic's liability.
Revenue was $16.3 billion for the six months to June 30, 2026, against $1.2 billion a year earlier; the operating loss was $693 million and the net loss $80.7 billion after $72.4 billion of fair-value charges on convertible instruments.
Liberty's cyber head called the language "a signal" and asked how to model build-up when one model sits under thousands of insureds; WTW said not every AI exposure can be assumed insurable.

What the filing says

The insurance language appears under a risk factor on liability from harms associated with the company's models. The filing says such harms, "whether arising from misuse by third parties, from our models' own behavior, or from their deployment in high-risk or sensitive contexts," may expose Anthropic to litigation under legal theories that remain unsettled for AI, and that "our contractual and insurance protections may prove insufficient."

It then sets out three limits. First, the policies Anthropic holds "may not provide sufficient coverage to adequately mitigate the potential liabilities we face, especially where any one incident, or a group of incidents, causes disproportionate harm." Second, "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." Third, "the insurance market for AI-specific risks is nascent and evolving," and the company "may be unable to obtain coverage for certain AI-related liabilities on commercially reasonable terms, or at all."

The same risk factor says the models "may behave in ways that no user or developer intended," so claims arising from those actions "could result in substantial liability." A separate factor on cybersecurity says the company maintains cover for breach costs but that claims it submits may not be "covered, paid in a timely manner, or paid at all." A third says its insurance may not compensate it for business interruption if the Amazon, Google or Microsoft infrastructure that hosts its models fails.

Exposure named in the S-1	What the filing says about cover
Harms from model behavior or misuse	Policies may not suffice for a disproportionate incident; a catastrophic incident may exceed any insurer's or the industry's capacity
AI-specific liabilities	Market is "nascent and evolving"; cover may be unavailable on commercially reasonable terms, or at all
Copyright claims over training data	Insurance "may not fully cover such claims," as in Bartz
Indemnities to evaluators and customers	Certain agreements carry no cap on Anthropic's liability
Cybersecurity and data incidents	Claims may be rejected, paid late or not paid; cover may become unavailable
Infrastructure outages	Insurance may not compensate for business interruption at cloud providers
Directors and officers	Public-company rules expected to make D&O cover more expensive, with reduced limits possible

Source: Anthropic Form S-1, risk factors, filed October 26, 2026. Paraphrased except where quoted.

The Bartz line

The filing gives one worked example of a claim its cover did not absorb. The consolidated statement of operations carries a $1.5 billion "Bartz settlement" line in 2025 operating expenses, the class action by authors over books used in training. In the risk factors, Anthropic says it may settle disputes "even where we have meritorious claims or defenses," as it did in Bartz, and that "as in Bartz, insurance may not fully cover such claims, provide sufficient payments to cover all the costs to resolve one or more such claims, or continue to be available on terms acceptable to us."

The company also says adverse rulings against other AI developers "regardless of the specific facts" could set precedents that limit its ability to train or deploy models, and that copyright regimes outside the United States generally do not recognize fair use. For an underwriter, the settlement is a data point on severity for a loss class that did not exist five years ago, paid by a company whose operating expenses in 2025 totaled $12.6 billion.

Against that, the filing reports revenue of $16.3 billion for the first half of 2026, from $1.2 billion in the same period of 2025 and $4.6 billion for all of 2025. Compute and infrastructure cost $8.9 billion in the half. The net loss of $80.7 billion is mostly a $72.4 billion change in the fair value of convertible notes, derivative liabilities and SAFEs, a non-cash charge that rises with the company's valuation.

Uncapped indemnities

Two contractual disclosures matter to anyone pricing the liability. The S-1 says "certain of our agreements with external evaluators and testing vendors require us to defend and indemnify the evaluator, without a cap on our liability," against third-party claims including intellectual property infringement arising from authorized testing. It also says some customer agreements require Anthropic to indemnify the counterparty, "without a cap on our liability," against claims that its models or their outputs infringe third-party rights.

Those terms sit alongside the terms-of-service disclaimers the company uses to limit its exposure to users, which the filing says "may not be upheld, or may not apply to certain claims." Together they describe a liability that is contractually open-ended at one end and legally untested at the other.

The filing also lists where the models are used. As of September 30, customers paying $100,000 or more a year included every Fortune 10 company, more than half of the Fortune 500 and more than 30% of the Global 2000. Among the production workloads it names is "underwriting insurance policies." Allianz signed a global partnership with Anthropic in January, Insurance Business reported.

The market's read

Rian Wickens, head of cyber and IT liability in Liberty's professional and financial risks business in Asia Pacific, said the passage was "more than a standard disclaimer. It's a signal." He posed three questions for carriers: "Where is AI exposure already hiding in our cyber, tech E&O, PI, D&O and casualty wordings without being named?", "How do we model the build-up of risk when one AI model sits underneath thousands of our insureds?" and "If the worst-case losses can't be insured, what's the backstop?"

"We've been here before on silent risk in insurance policies with terrorism, pandemic and cyber," Wickens said. "AI won't be any different."

Andrew Hill, global FINEX AI risk strategic leader at WTW, said "the fundamentals of insurance as a risk transfer strategy for AI risk remain robust," but that as understanding of the risk deepens, "particularly its potential for systemic and aggregated losses, it cannot be assumed that all AI exposures will be insurable."

A Gallagher survey cited in the same coverage found one in five insurance professionals had a client suffer an AI-related loss or claim in the past year, and just over half of those were fully covered. Daniel Elechiguerra Batlle, cofounder of Harbor, an insurer of AI agents, said the risk "is just going to become part of our lives, and therefore we need to learn how to live with that," adding that "guardrails and the controls are going to improve."

The aggregation question Wickens raised is the one the S-1 answers least. The filing describes concentration from the developer's side, a limited number of customers and distribution channels driving much of its revenue, but not the number of insureds whose operations depend on a single model release. That figure sits in carriers' own books, in policies that in most cases do not name the exposure. The filing's risk factors run to 94 pages, and the offering has no date.