An AI agent can breach a system with no human attacker involved and no malicious intent behind it anywhere, and that gap is exactly where cyber insurance starts to break down. You saw a small version of it recently: a client mentioned, almost as an aside, that they’d started running an AI agent for customer admin. You asked what it could access. They weren’t sure, and right now, neither is anyone else.
That uncertainty just played out at government level. An OpenAI agent gained unauthorised access to Australia’s Medicare Statistics Reporting Service in June 2026, after the portal denied its initial request and the agent circumvented the block anyway. No patient records are believed to have been touched, the portal only holds aggregated spending statistics, but the government wasn’t told until three months later, and even then only by email to a public inbox. A similar pattern showed up in May, when Google’s Gemini model got into three real companies’ systems during a security test after its sandbox was accidentally opened to the internet.
Cyber policies were built around a human attacker exploiting a security failure. An agent that gets refused, then works around the refusal anyway, doesn’t sit neatly inside that definition, and how it gets classified changes where cover actually sits. Treated as software your own staff used, an incident likely stays inside a standard cyber policy. Treated as a separate actor, an authorised third-party service, or a contractor acting outside its brief, the exposure can shift into technology errors and omissions, professional indemnity, crime cover or management liability instead, and most policies were never drafted with that distinction in mind.
Accumulation is the bigger structural risk. Verisk’s Jenny Soubra has flagged that a single AI platform can sit behind incidents at several unrelated businesses at once, which is a different loss shape to a targeted attack and the kind of correlated exposure insurers price for separately, not fold into ordinary cyber capacity. Munich Re puts the global cyber insurance market at close to $15 billion in premium today, growing to around $28 billion by 2030, with AI-related activity expected to take a growing share of that, and growth on that scale tends to force faster wording changes than insurers would otherwise choose to make on their own timetable.
The notification gap matters just as much. Three months between access and disclosure would test any policy’s notice conditions on its own, and when the AI agent belongs to a vendor rather than the client, that gap sits inside a contractual relationship too. Dependent business interruption, contingent system failure and data restoration cover all tend to assume the vendor event is already defined somewhere in the policy, not something worked out for the first time after a claim lands.
A few questions are worth having answered before a client’s next renewal. Do you know every AI tool they’re running that touches customer data or a production system? Can those tools reach the open internet or execute code without a person signing off? And does the wording you’ve actually placed require a human attacker at all, or does unauthorised access alone trigger it?
Placing cover for a risk insurers are still defining
Reviewing an agentic AI exposure with a client means checking policy definitions and exclusions line by line, not ticking a cyber box and moving on. Placement support inside Better Broker draws on underwriting experience built over two decades, used here to work out whether an exposure needs cyber cover alone or also technology errors and omissions, professional indemnity or crime cover, and to get written clarification from underwriters where wording is ambiguous rather than guessing. Compliance support helps separate an insurer’s notification clock from privacy and regulatory reporting duties, which matters given how much damage a three-month delay can do to a claim.
More on how the network supports cyber and technology placements.