Probably, because a staff member pasting client data into a chat tool looks like an ordinary privacy event rather than an AI exclusion.
Short answer: Most likely yes, though it depends entirely on your wording. Insurers have been treating AI-enabled incidents as ordinary cyber perils such as unauthorised access, privacy breach and social engineering fraud. The AI exclusions that appeared in 2026 are mostly in other lines of insurance, and they target losses caused by your own AI systems, not losses caused by a staff member using someone else's chat tool badly.
Key takeaways
Start here, because almost every article on this subject asserts the opposite.
Marsh's Global Insurance Market Index recorded global cyber insurance rates down 4 per cent in the second quarter of 2026, the twelfth consecutive quarter of decline, following a 5 per cent fall in the first quarter. Broad commercial insurance rates fell too, driven by plentiful capacity and strong competition between insurers.
So if someone tells you AI attacks are pushing cyber premiums up across the market, the published data does not support it. What is true is subtler and more useful: market averages and individual renewals have come apart. Plenty of businesses are seeing higher quotes inside a softening market, because underwriters are pricing the individual risk rather than the category. Your controls, your claims history and your sector move your number more than the cycle does.
That is good news for a well run small firm and bad news for a poorly documented one.
This is the distinction worth understanding, because it decides almost everything.
An attacker's AI. Someone uses a cloned voice to impersonate your director, or generates a convincing invoice, or writes a flawless phishing email at scale. Cyber policies have generally continued to treat this as what it is: fraud and unauthorised access, dressed in better language. Analysis of the 2026 market found cyber remained the most stable source of coverage for AI-related risk, with some carriers issuing endorsements specifically confirming that deepfake and AI-driven attacks are covered.
Your own AI. You deploy an autonomous agent that takes an action, or a model produces an output that causes loss, or your AI governance is found wanting. This is where the gaps are opening, and it is where new exclusions and conditions are appearing.
An employee pasting a client's bank details into a consumer chat tool sits squarely in neither of those. It is an unauthorised disclosure of personal information by an insider, which is one of the oldest covered events there is. That is why the honest answer to the headline question is "probably yes, and for boring reasons".
Read carefully here, because the coverage of this topic conflates lines of insurance constantly.
The Insurance Services Office published a generative AI exclusion in January 2026 for commercial general liability policies, excluding bodily injury, property damage and personal or advertising injury attributable to generative AI. That is not your cyber policy. The broader fragmentation and tightening has been concentrated in management liability lines such as directors and officers cover, again not cyber.
Some insurers have added AI security riders that ask for evidence of red teaming and documented risk assessment before extending certain cover. Those are conditions on writing the risk rather than exclusions of it.
None of this means your cyber policy is safe by default. It means the thing to look for is not an "AI exclusion" heading. It is the ordinary machinery: your definition of a privacy breach, whether unauthorised disclosure by an employee is covered, your sublimits for regulatory response costs, and any condition requiring you to maintain stated controls.
The controls that appear on Australian small business cyber applications are consistent and short: multi-factor authentication on email and remote access, endpoint protection, tested backups, patching, email filtering, staff awareness training, and a written incident response plan.
Notice what is not on that list. Data loss prevention is not a standard named requirement at small business level. It appears on enterprise questionnaires, not on a form for a firm of twelve people.
Notice also what has changed in tone. Underwriters increasingly want evidence rather than a tick. Saying you have multi-factor authentication is different from showing where it is enforced, and a written incident response plan means a document, not an intention.
You will find plenty of claims that specific controls cut premiums by fifteen, twenty five, even forty per cent. Follow those figures back and they lead to broker marketing, managed service provider blogs and vendor pages, not to a carrier rate filing or a regulator.
The defensible version is that controls affect eligibility, terms, sublimits and price, sometimes substantially, and that undocumented controls are treated as absent. That is worth acting on. A precise percentage is not something anyone can promise you.
Plainly, so there is no confusion: Prinn does not lower your premium, is not recognised by any insurer, will not get you a discount, and does not make you compliant with anything. Any vendor telling you otherwise is guessing on your behalf.
What it does is narrower and real. Cyber applications ask how you handle sensitive data and who can access it. Most small firms answer those questions with intentions. Prinn checks a prompt before it is sent, and if it finds an API key, a card number or a passport number it shows the person what it found and offers three choices: redact it, stop, or send it as is. The person decides, and it never blocks anyone. Prompt content never leaves the machine, including to us.
The part that matters after an incident is the record. An owner sees the category, the tool, the date, the device and the choice made, without reading anyone's prompts. If you ever have to tell an insurer or a regulator what was exposed and when, the difference between a documented answer and a shrug is the whole conversation.
It is reliable on credentials, financial details and identity numbers. Names, home addresses and dates of birth in ordinary prose are harder, and it is still improving at those.
Ask your broker four questions, in writing.
Does the policy respond to unauthorised disclosure of personal information by our own employee, including where the disclosure was to a third party service? Does it contain any exclusion, endorsement or condition referring to artificial intelligence, and if so which line does it sit in? What controls are we warranting we have in place, and would we survive being asked for evidence? And what are the sublimits on regulatory response and notification costs, given notifying affected people is usually the first real cost.
Get the answers before you need them. The worst time to learn how your policy defines a privacy breach is while you are drafting the notification.