Last week, the company behind one of the world's most powerful AI models told the SEC what the insurance industry has been saying in private for two years. Anthropic, the maker of Claude, filed its Form S-1 on October 26 ahead of a planned Nasdaq listing under the ticker ANT. Its 94 pages of risk factors say the harm from a catastrophic AI incident could exceed what any single insurer — or possibly the industry as a whole — has the capacity or expertise to underwrite. It calls the market for AI-specific cover "nascent and evolving." It may not be able to buy protection for some AI liabilities on commercially reasonable terms. Or at all.
This is not a hypothetical. Anthropic already knows what an uninsured AI loss feels like. In September 2025 it agreed to pay $1.5 billion to settle the Bartz copyright class action — a third of its entire 2025 revenue — and its insurance did not fully cover it. As of June 30, 2026, it still carried $1.2 billion of the settlement as a liability. The filing also discloses uncapped indemnities under some agreements with outside model testers. The largest AI developer to go public is telling investors, under oath, that some of its risk cannot be transferred at any price the market will offer.
What it means for carriers
First, the protection gap is no longer a conference warning. It is a disclosure. When a frontier developer tells its regulator the industry may not absorb its worst-case risk, that is not boilerplate. Rian Wickens, head of cyber and IT liability at Liberty in Asia Pacific, called it "more than a standard disclaimer... it's a signal." Andrew Hill, WTW's global FINEX AI risk leader, added the sentence carriers should underline: "it cannot be assumed that all AI exposures will be insurable."
Second, the S-1 puts a ceiling on the market split this series has tracked. In October, specialists began writing affirmative AI cover while large carriers excluded generative AI from general liability. That split was rational. But affirmative limits run $2 million to $50 million, and a catastrophe is a different order of magnitude. Both camps now face the same question from opposite ends: what does the industry do when the risk it is asked to underwrite is bigger than the industry?
Third, the real story is accumulation. Morningstar DBRS's Marcos Alvarez noted that a widespread model failure could hit many customers simultaneously, generating claims across different policies while interrupting insurers' own operations. One model, thousands of insureds, one failure mode. This is catastrophe risk wearing a liability costume, and most carriers still underwrite it one policy at a time, as if the losses were independent. They are not. A Gallagher survey found one in five insurance professionals had a client suffer an AI-related loss last year; only just over half were fully covered. The gap is already showing up in ordinary claims.
From the operator's chair
One: name your catastrophic position, line by line, before your next renewal cycle names it for you. Decide explicitly, line by line, what AI risk is in, out, or priced — with one named owner across the firm. Silent AI exposure unpriced in casualty and cyber wordings is how a protection gap becomes a balance-sheet event.
Two: underwrite the controls, not the model. Carriers already ask in supplemental applications whether a company keeps an inventory of its AI tools and use cases, and who owns each deployment. Make that discipline your own standard first: inventory, named owner per deployment, guardrails, human oversight, supply-chain mapping. A carrier that can show that discipline in its own operations can underwrite it in others.
Three: give every agent a mandate it can lose. On November 4, Singapore's parliament debated exactly that: a national digital identity for AI agents — an open registry where any party could check an agent's identity and mandate, with revocation built in. The government backs the idea, and its GovTech agency is already building an internal registry for 150,000 public officers. The verifiable mandate this series argued for is becoming state infrastructure. An agent with no named owner and no revocable mandate is uninsurable by definition.
Four: price model concentration the way you price cat accumulation. One foundation model inside a thousand insureds is a single point of failure with a thousand claim paths. Aggregate it, stress it, and charge for it — or stop pretending the policies are independent.
The insurance industry was built to absorb the losses nobody else could. Anthropic just told the SEC there is now a loss it may not be able to absorb. That is not a reason to stop underwriting AI risk. It is a reason to underwrite it on purpose — named owners, revocable mandates, limits that mean something — before the market finds the gap the expensive way.
— Krishna