I recently asked Christopher Keegan (Brown & Brown) to share his perspective on where the cyber insurance market is heading. Everyone knows Chris is a straight shooter, and his answers were incredibly candid.
We covered:
- Why board-level AI governance matters more than technical sandboxes
- Where AI is actually showing up in claims today
- Why he doesn’t expect a large systemic event in the next year
The Rate Environment
Q: Cyber rates have fallen since 2023, but the decline is slowing, according to Marsh. Are we at the bottom? And what would turn the market around — a systemic event?
Christopher Keegan: Agree with that — for the market to turn, losses have to go up, and we’re not seeing that. I try to track where losses are, but the information is inconsistent. Carriers continue to enter the market and competitively need to build a book. Some lowered rates writing programs without sufficient premium, and have been impacted by claims. Some of those markets have gone out of business, but there’s still plenty of capacity, and more companies are coming in. As long as that continues, and the loss ratio stays around 70%, the market will stay competitive.
How Underwriters Actually Evaluate Risk
Q: Are carriers still testing an insured’s environment with outside controls validation — underwriting calls — or are they mostly relying on questionnaires and self-attestation?
Christopher Keegan: It’s different for different parts of the market. Very small SMB accounts answer four or five questions — some markets ask no questions at all on renewals, just gather information utilizing an online scan platform. Middle-market applications are more in-depth, anywhere up to 40 or 50 questions. For large accounts, we often dispense with an application entirely and provide information on a call instead. The market does not utilize internal network scans much at all — the idea of putting a little box on the system is not something that insureds are open to.
Q: Is underwriting really moving from a controls checklist to an exposure-driven view — and what does that mean in practice?
Christopher Keegan: They are. But carriers still run through the controls — they still want to know if you’ve got MFA, PAM, backups, encryption, all of the critical protections.
Q: What’s the frequency and severity mix you’re seeing — ransomware, data theft, funds transfer?
Christopher Keegan: For smaller companies, it’s really about 50/50 split between ransomware and funds-transfer fraud. As you get into larger companies, it’s more purely ransomware.
Where AI Actually Shows Up in Claims Today
Q: Where has AI clearly been a factor in claims you’ve seen?
Christopher Keegan: The only claims where AI is really involved that we’re seeing are intellectual property claims — companies training their models on information they shouldn’t be. The courts will tell us whether that’s right or not. I’m sure ransomware attacks now are being generated by AI, and the lateral movement within organizations once attackers get in is coordinated by AI — but having AI identified as a cause in a specific claim is still rare.
Q: What about deepfakes — has that come up on any claims?
Christopher Keegan: On small accounts, I wouldn’t say it’s so much deepfakes — it’s more social engineering. You don’t need to get that sophisticated to take down a small company.
AI Coverage: Which Policy Actually Responds
Q: With AI-related incidents, does coverage really only fall under cyber, or does it show up in tech E&O and other lines too?
Christopher Keegan: Cover for AI could implicate tech E&O as well — if a company is providing an AI service, that would be covered under tech. If a law firm is using AI to do its research and it has an error in the research, that’s going to be covered under their lawyers’ professional liability policy. ISO has come out with AI exclusions, but each carrier chooses whether it wishes to use them — and what we’re seeing is that they’re barely being used at all, and not in the cyber market. Where I have seen them used is in the casualty market — on employment practices liability, for example, where a company’s AI-driven resume screening raises discrimination concerns.
Governance: What Underwriters Actually Want to See
Q: How are underwriters understanding and managing AI risk day to day? Do they want an AI inventory, agent permissions, human approval on high-risk actions, monitoring?
Christopher Keegan: They know they can’t dig deeply into every piece of agentic AI. What they’re looking for is governance and compliance procedures — board oversight, roles and responsibilities clearly set, an AI governance committee that’s multidisciplinary, and enough training in place to limit shadow AI use.
Q: Do identity controls that contribute to good governance make it any easier for underwriters to assess AI risk specifically?
Christopher Keegan: Not often. It’s all still really very high level.
Q: Is the insurance community open to something like a national AI testing standard — to avoid another incident like OpenAI–Hugging Face — or is the attitude more “it’s early days, let them innovate, we’ll figure it out”?
Christopher Keegan: It’s unlikely that insurers are going to impose standards on AI. Apart from requiring governance — if you don’t have good AI governance in place and you’re letting people run amok within your organization, carriers will punish you with limited terms and higher premiums. In their view, there has to be some level of control over the people. So you need an inventory of your AI systems, and controls over the people using them.
The Nvidia Sentry Announcement
Q: Nvidia just announced Sentry — a hardware-based sandbox where each agent starts with predefined permissions, and a separate chip monitors and enforces policy outside the agent’s own control, independent of whether the agent itself is compromised. Does independent, hardware-level monitoring like that change how underwriters would look at an agent deployment?
Christopher Keegan: That’s a common refrain from IT Security providers — the number of times I’ve heard the phrase “silver bullet” over my lifetime in cyber insurance. We have an example of a large bank — they said, “we don’t have a need for cyber insurance, we’re locked down,” and a week later it was a headline. Solutions such as Sentry need to be tested. But even then, it’s probably too narrow a solution to really make a difference to the underwriting by itself. It may become a standard eventually, but it usually takes the market a while to coalesce around something in order to get recognition as an expected protection. Look at MFA — it took two or three years to become a real standard, and even a year or two after it was clear you should have it, some markets would still waive the requirement just to get the premium booked.
Shadow AI, Third-Party Risk, and Board Oversight
Q: Do applications or renewal conversations ask about shadow AI specifically — whether the insured can see which AI tools are in use, or whether they block unapproved ones? Would a documented AI-use policy with technical enforcement affect terms?
Christopher Keegan: Not specifically. Underwriters ask about training, and the expectation is that if you’ve got training in place, training should include a shadow AI policy, and you’ll be requiring compliance with it. Underwriters expect you to have an inventory and enforce the rules about what’s in it. However, carriers are not asking specifics, like how many shadow AI instances you found in the last year — I really haven’t seen that question come up.
Q: For heavily regulated environments — PCI DSS, HIPAA — where agents might be able to touch sensitive systems, is there additional scrutiny?
Christopher Keegan: I haven’t had that question come up recently.
Q: If a company can affirmatively say they have governance, an inventory, and a training program in place, does that mean you’re not going to dig further into exactly how they monitor or test?
Christopher Keegan: Yes, at the moment, that’s right.
Q: Breaches now happen in seconds instead of minutes or hours. Would you credit continuous control validation over an annual point-in-time attestation?
Christopher Keegan: That’s definitely happening. I think the attacks are going to be so quick, and access to the crown jewels is going to happen fast. However, currently people are still making the decisions about whether to cut off all or part of a network. That’s still a human decision, but I’m not sure how long that situation will last. At some point, I think there has to be some level of AI that comes in and makes those decisions on behalf of people. AI will be attacking and defending. That may happen soon.
Q: Do you ask about dependence on third-party software — essentially an SBOM? Is that something insurers are asking about now?
Christopher Keegan: Yes. Historically, it’s been a question.
Q: Would board-level evidence of AI oversight matter in cyber or D&O underwriting?
Christopher Keegan: Yes. Carriers really want the board to be involved — because once the board is involved, everybody takes it seriously. I call it “tone at the top.” Making sure people know it’s important because it’s a board-level issue.
Q: Is there anything you think the carrier — or the board — should be asking about these risks that isn’t being asked right now?
Christopher Keegan: I do think you’ve got to corral staff — get people to do what you want them to do. Every company should be trying to get people on track to: understand the risk, do the inventory, do the governance, understand it, and push people in a certain direction. Tell them for example: don’t put your board report into public ChatGPT — that kind of thing gets you fired.
Looking Ahead: The Next 12 Months
Q: Twelve months from now, do you expect anything to be materially different — coverage, regulation, or the attacks themselves?
Christopher Keegan: It’s not clear where the systemic attack is going to come from. Remember MOVEit — I’d call that a systemic event. It was a supply-chain issue: a lot of companies had it, a lot of companies had the vulnerability, a lot of companies didn’t fix it, and there weren’t enough threat actors out there to actually work through all of them. I think AI-driven attacks are going to be similar — the vulnerabilities will be out there, but as with the CrowdStrike failure and other similar situations, a lot of those vulnerabilities are going to get closed off before there are adverse consequences.
Companies are better at recovering — five years ago, something like 85% of ransoms were being paid; now it’s under 20%, because companies can recover from immutable backups. On the bigger, systemic events — taking down infrastructure, water, 911 systems — my view is that’s more likely to be a nation-state attack, not ransomware actors; usually, they’re not doing that for money. So a year out, I don’t think we’ll see one of those big events. I think companies, with new tools from IT security companies like CrowdStrike and others, are going to be able to manage the risk — there will still be individual events, but not large systemic ones.
— Barnaby Page, Page AI Studio