Insurers divided over who pays when AI agents cause harm

Companies deploying AI agents may not be covered if those systems cause damage, according to a report published by Dark Reading on 22 October 2026. The publication says the cyber insurance market has split: some insurers are building specialised coverage for agentic AI risk, while major carriers are quietly excluding it from their policies. The result, Dark Reading says, is that enterprises can be left exposed without knowing it.

The detail that matters most for security teams is the word "quietly". An exclusion is a clause that removes a category of loss from what the policy will pay out on. If the exclusion is new, narrow or buried in renewal paperwork, a buyer who has not read the wording closely may assume the cover they had last year still applies. Dark Reading's framing is that this gap between what firms believe they have bought and what they have actually bought is the live problem, not any single incident.

The divide also tells you something about how insurers see the risk. Writing a specialised agentic AI product means an insurer believes the exposure can be defined, priced and limited. Excluding it means the opposite: that the losses are too hard to model, or too correlated across many customers, to underwrite at a sensible premium. Both positions are rational responses to the same uncertainty, and according to Dark Reading both are being taken in the same market at the same time.

Agentic AI is the term for systems that do not just answer questions but take actions, such as calling tools, moving data or making changes in other systems on a user's behalf. That autonomy is what makes the insurance question awkward. A traditional cyber policy is built around an intruder breaching a network. An agent that has been given legitimate credentials and then does something harmful is not obviously a breach, and may not be obviously an error or omission either. Where a loss lands between policy categories, it risks landing nowhere.

For buyers, the practical takeaway from the report is to read the renewal wording rather than rely on the broker's summary, and to ask directly whether losses caused by autonomous or agent-driven systems are covered, excluded, or simply unaddressed. Silence in a policy is not the same as cover. Firms that have deployed agents since their last renewal may also find that the risk they disclosed at underwriting no longer matches the risk they are running.

Dark Reading does not report a specific claim being denied, and no insurer is named in the summary of its coverage. The article describes a market in motion rather than a settled position, which means the terms available to a given buyer are likely to depend on their carrier, their sector and the timing of their renewal. Security teams that cannot answer where their agent-related losses would sit should assume the answer is currently with them.