Most cyber insurance policies in force today were underwritten for a world where AI is something attackers use against you — phishing content, deepfake fraud, malware — not something your own organization deploys with the authority to take autonomous action. That's the gap: coverage exists for AI-enabled attacks; coverage for losses your own agent causes is thin, fragmented, and — as of August 2026 — still being built.
This post lays out that gap, what changed in policy language in early 2026, and the questions worth taking to your broker before your next renewal. It is not AI agent governance for insurance — that post covers how insurance companies govern the AI agents they deploy in underwriting and claims; this one covers buying your own coverage for what your organization's agents do.
The gap in one paragraph
The core problem is a mismatch between what cyber policies were written to cover and what agentic AI actually does: policies largely anticipate AI as an attack vector used against a policyholder, not as an autonomous system a policyholder deployed that can independently cause a loss — a bad tool call, an unauthorized transaction, a data exposure with no external attacker involved at all. An agent that takes a harmful action on its own, with no human in the loop and no external attacker, doesn't cleanly fit the incident categories most cyber policies were drafted around.
What changed in January 2026: the ISO exclusion in CGL policies
In January 2026, ISO introduced a new generative-AI exclusion into commercial general liability (CGL) policies, adding to the fragmentation of coverage across policy types. CGL policies were never the primary home for cyber risk, but the exclusion matters as a signal of direction: standard-setting insurance bodies are actively drawing new boundaries around AI-related losses rather than leaving them to fall wherever existing silent-cyber language happens to land. Exactly what a given exclusion carves out varies by how it's incorporated into a specific policy — that's a question for your broker and your own policy language, not something this post can generalize from the fact of the exclusion's existence.
Coverage fragmentation across policy types
No single policy type is currently the clear home for agent-caused losses — cyber, CGL, tech E&O, and D&O policies each cover adjacent pieces, and the boundaries between them are exactly where the ISO exclusion and similar moves are actively being drawn.
| Policy type | Traditionally covers | Where an agent-caused loss might land |
|---|---|---|
| Cyber | Data breaches, business interruption from a security incident | If the agent's action produced a data exposure or system outage |
| CGL | Third-party bodily injury and property damage claims | Now carries a generative-AI exclusion, introduced January 2026, that may narrow this path |
| Tech E&O | Errors and omissions in a product or professional service you provide | If the agent was part of a product or service sold to a customer |
| D&O | Claims against directors and officers for governance failures | If a loss is framed as inadequate oversight of an AI deployment decision |
A loss caused by an autonomous agent might implicate more than one of these at once, with each insurer pointing at another's policy as the more appropriate home. That fragmentation is the practical reason a broker conversation, not a single policy read, is the right first step — the answer to "what covers this" often depends on which policy gets asked first, and on how the loss gets framed in the initial claim notice.
What underwriters are starting to ask for at renewal
Cyber insurers increasingly evaluate AI governance maturity, documented risk assessments, and tool inventories as part of underwriting and renewal — not just the technical security controls a traditional cyber questionnaire has always asked about. That shift means the underwriting conversation is starting to resemble a governance audit as much as a security audit: can you name the agents you've deployed, what they're authorized to do, and what evidence you have that you're managing the risk they carry. Insurers asking these questions are, in effect, pricing risk based on governance maturity the same way they've long priced it based on patch cadence and access controls — the AI-specific version of underwriting is following the same logic, applied to a newer risk category.
The governance artifacts that map to those underwriting questions
The artifacts underwriters are starting to ask for line up closely with artifacts a mature agent governance program already produces for other reasons. A documented view of your agent governance maturity — which agents are deployed, at what stage of control, and what's authorized — answers the "what have you deployed" question directly. Evidence from a documented red-teaming program answers the "have you tested for failure modes" question. And an incident readiness plan that covers agent-caused events, not just external attacks, answers the "what happens when something goes wrong" question underwriters increasingly want addressed before, not after, a claim.
| Underwriting question | Governance artifact that answers it |
|---|---|
| What AI agents do you run, and what are they authorized to do? | Agent inventory / governance maturity documentation |
| Have you tested for failure modes before deployment? | Red-teaming evidence, documented risk assessments |
| What happens when an agent causes an incident? | Incident readiness plan covering agent-caused events |
| How do you know what an agent actually did? | Audit records attributable to a specific agent and action |
None of this guarantees favorable terms; it's the evidence base that lets an underwriter price your risk instead of declining to, which is a precondition for coverage existing at all in a market that's still forming. Organizations without this documentation aren't necessarily uninsurable, but they're negotiating from a weaker position — an underwriter with less evidence to price against tends to default to narrower terms or a higher premium, not more generous ones.
Questions to take to your broker before the next renewal
This is where a broker conversation belongs, not a conclusion this post can draw for your specific policy. Worth asking directly: does our current cyber policy cover a loss caused by our own deployed AI agent with no external attacker involved, or only AI-enabled attacks against us? Has our CGL, tech E&O, or D&O coverage been amended with any generative-AI exclusion language since January 2026, and if so, what exactly does it carve out? What documentation — agent inventory, red-teaming evidence, governance program maturity — would improve our terms or reduce a coverage dispute at claim time? And if a loss spans more than one policy type, which insurer is the first point of contact? None of these have a generic answer; they're a starting list for a conversation with your broker and, where the exposure is material, your counsel — not a substitute for either.
Where this is heading
As of August 2026, the market for AI-agent-specific cyber coverage is actively forming rather than settled: exclusion language is being introduced, underwriting questions are shifting toward governance maturity, and no standard, stable policy language for autonomous-agent losses exists across the industry yet. That's a reason to start the broker conversation now rather than wait for the market to stabilize — organizations that can already produce governance evidence when asked will be better positioned as underwriting standards solidify than the ones scrambling to produce it retroactively at a renewal deadline. A governance program built for its own sake — identity, guardrails, audit evidence — happens to be the same evidence base insurers are starting to ask for.
Common questions
Does my existing cyber insurance cover losses caused by my own AI agent? It depends on your specific policy, and this post can't answer that for you — that's a question for your broker. What's broadly true as of August 2026 is that most cyber policies were written with AI-enabled attacks in mind, not autonomous losses caused by an organization's own deployed agent, and coverage for the latter is fragmented and still forming.
What is the ISO generative-AI exclusion? In January 2026, ISO introduced a new generative-AI exclusion into commercial general liability (CGL) policies. What exactly it excludes depends on how it's incorporated into a specific policy — confirm the exact language with your broker rather than assuming a generic scope.
Is this the same as the "AI agent governance for insurance" post on this site? No. That post is about how insurance companies govern the AI agents they deploy internally, in underwriting and claims. This post is about how an organization, in any industry, buys coverage for the losses its own AI agents might cause.
What should I bring to my insurance broker before renewal? An inventory of the agents you've deployed and what they're authorized to do, evidence of red-teaming or documented risk assessment, and an incident-readiness plan that covers agent-caused events specifically, not just external attacks. These are the artifacts underwriters are increasingly asking for at renewal.
Should I wait for the AI-agent insurance market to mature before addressing this? No. As of August 2026 the market is actively forming, not finished, which means waiting doesn't buy certainty — it just delays the broker conversation. Building the governance evidence base described above is worth doing regardless of how quickly the insurance market standardizes around it, since the same documentation supports regulatory compliance and internal risk management independent of any insurance outcome.