Estimated reading time: 7 minutes
Agentic AI is reshaping how underwriters and claims teams think about cyber risk. The disruption arrives at machine speed. AI keeps inventing and changing at a pace the market rarely sees. Insurance touches every industry. Every industry is now embracing AI. That pushes risk and liability to the forefront. The latest Cyber Insurance News podcast episode digs into that pressure. We recorded the panel live at last week’s Scout InsurTech conference in Columbus, Ohio. The episode carries a sharp title: “The Authorization Gap: Cyber Insurance in the Age of Agentic AI.” The panel dug into the complex problem of insuring agentic AI. Who carries the blame when an agent fails? Where do liability and responsibility land? Which controls hold? Who implements them?
The conversation drew four perspectives from across the cyber insurance value chain. Each panelist sees the same risk from a different seat.
Who Is on the Panel
Julia Garcia-Trombley leads US & Canada for CertX. She works at the intersection of control verification and underwriting confidence. Jeremy Epstein is CEO of Mayflower Specialty. He focuses on pricing the risk. Rich Gatz is Head of Cyber Claims at Arch Insurance. He pays for the risk after a loss. Tristan Morris is CEO of SplitSecure. He builds the control layer between delegated authority and execution.
Get The Podcast At These Spots
YouTube
Spotify
Amazon
Apple
What Agentic AI Means for Insurers
The panel started by defining agentic AI as a system that acts without human involvement. It can call APIs, run code, or use other tools on its own. If it only gives advice to a person, it is not considered agentic. This difference affects how insurance coverage applies.
The Risk Lives in the Permissions
A main theme of the session was that the real risk comes from the access the AI model has. Controls should be placed on the surrounding systems, not just on the AI itself.
The panel used a helpful analogy: Imagine a smart but unreliable intern with admin access to a key system. If the intern falls for a phishing attack and lets in an attacker, the responsibility lies with the person who gave them access. The same idea applies to AI agents; untrusted actors should never have unlimited permissions.
This changes how underwriters approach the issue. They need to consider what an AI agent could do in the worst-case scenario and understand the maximum possible damage if the model fails or is misused. The scope of permissions becomes the main way to measure risk.
Is Cyber Insurance the Right Home?
Carriers have different views on this issue. Most current cyber policies are broad, and the market has remained soft for years. Premiums dropped after a sharp increase during the pandemic, so many carriers now provide wide coverage at a low price.
Two foundational triggers sit in a standard cyber policy. One is a network security breach. The other is a data privacy breach. An AI-driven breach usually falls under those triggers already. From a claims view, the cause of the breach often does not change the payout.
Some panelists suggested looking at AI risk more broadly. They said AI risk goes beyond cyber and affects errors and omissions, directors and officers, and employment practices liability. Because AI output is unpredictable, it creates new risks that older policies did not expect. Some carriers now exclude AI, while others offer specific AI coverage. A few have even arranged special reinsurance for AI liability.
Controls That Actually Work
Certification is one way to address the confusion. A control either works or it does not. Assessors can test systems against set boundaries and use test suites to make sure the limits are effective. If the system breaks its guardrails, it loses certification.
Risk can never be completely eliminated, but good governance can reduce it as much as possible. We haven’t removed all risk and made cars, planes, or trains perfectly safe. The aim is to reach an acceptable level of risk, not perfection. Companies can still get real value from AI if they manage it properly.
There are many standards for AI and cybersecurity, and the field is still developing. About fifty standards exist, including the EU AI Act, ISO 42001, and ISO 27001, each with its own focus. No single framework works for every company.
The Threat Landscape Shifts
AI has not changed the basic ways systems are compromised, but it has made attacks more scalable. Social engineering is now much more convincing. Gone are the days of poorly written phishing emails; attackers can now quickly create polished, professional messages.
This increased scale works in other ways, too. Some law firms use AI to scan websites for privacy violations, focusing on cookie and tracking practices under old wiretapping laws. They then send out demand letters.
Most losses are still caused by human error, the flesh and blood that is behind most cyber claims. Strong and enforced controls can limit how much damage attackers can do. Defenders have to be right every time, but attackers only need to succeed once.
What the Industry Needs First
The panel ended with practical advice. Julia encouraged the industry to agree on definitions and keep up with regulations. She suggested doing a risk assessment before using AI. Jeremy called for more dedicated risk capital and better education for brokers. Rich recommended that any business using computers should have cyber insurance from a trusted carrier, and firms building AI products should also get technology E&O coverage.
Tristan suggested a simple approach: define what responsible AI use looks like instead of trying to list every possible error. Understand what actions the AI agent can take, make sure controls prevent it from gaining extra power, and keep a human in place to intercept sensitive actions. This helps keep the system manageable.
A Debate Bigger Than One Panel
The authorization gap and AI risk will outlast this conversation. Cyber Insurance News has tracked the risk, agentic AI, and otherwise and that won’t change. Security leaders are already slamming the brakes, per Apono’s 2026 agentic AI survey. Flashpoint logged a steep rise in agentic AI cybercrime. An OpenText study found risk management lagging behind AI adoption. Netskope flagged surging shadow AI and data leaks. GitGuardian counted millions of leaked credentials from AI-assisted coding. The panel is a snapshot. The story will keep moving. Much of the information you have today will perish under a weight only the future can bring.
FAQ – Agentic AI Cyber Insurance
Agentic AI is a system that takes action on its own. It can call APIs, run code, or operate other tools. It acts without a human in the loop.
Most standard cyber policies are broad today. They cover network security breaches and data privacy breaches. An AI-driven breach often falls under those triggers. Coverage still varies by carrier, so confirm the details with your broker.
Some carriers do not yet know how to price AI exposure. AI output is non-deterministic. That risk can extend beyond cyber into E&O and D&O. A few carriers treat it as a prohibited class for now.
The key control limits what an AI agent can do. Define its permissions clearly. Stop it from gaining more access on its own. Keep a human able to intercept sensitive actions.
Start with a risk assessment. Agree on definitions and check the relevant standards. Then build a plan and confirm your coverage. Ask your broker whether AI incidents are covered.
Related Cyber Insurance Posts
- CrowdStrike Brings Cyber Insurance Into Project QuiltWorks to Counter Frontier AI Risk
- The Human Factor: Why People Are Cyber’s Biggest Risk And Its Biggest Fix
- Cyber Liability Insurance: Key Takeaways From PLUS Cyber Symposium(Opens in a new browser tab)
- Non-Human Identity Sprawl Is a Cyber Liability Insurance Problem Now(Opens in a new browser tab)
- Artificial Intelligence Report: Only 44% Ready to Support Secure AI, Delinea Finds(Opens in a new browser tab)