The Next Cyber Insurance Claim May Start With Authorized AI

Jay Bavisi is the Founder, Chairman, and Group CEO of EC-Council, a global leader in AI governance and cybersecurity education.

getty​In the current insurance landscape, underwriters know how to ask about stolen credentials, ransomware, unauthorized access, data theft and the controls that reduce those risks. That work is still needed. In fact, the cyber insurance industry has spent decades getting better at measuring compromise.

In a world where IBM’s “Cost of a Data Breach Report 2026” put the global average breach cost at $4.99 million, up 12% from the prior year, AI agents create a different level of exposure because a system can create material loss while using valid credentials, following an approved workflow and operating inside the authority the business gave it.​

Gartner predicts that by 2028, one-third of interactions with generative AI services will use action models and autonomous agents to complete tasks. As more of those agents move into business workflows, cyber risk will include what authorized systems are allowed to do when they are operating normally.​

Two companies can use the same model and carry very different levels of risk. One may only read information and recommend a next step. Another may send emails, approve transactions, modify customer records, change identity policies or call another automated system. For insurers, model choice is only part of the assessment. The authority attached to the system determines how a mistake becomes a loss.​

Every AI agent should have an authority budget, meaning the maximum set of consequential actions it can take without human approval. That budget should be visible to the business, the security team and, where the use case is material, the insurer. Nonhuman identities, OAuth tokens, API keys, service accounts and privileged access management will become financially relevant because they determine where the agent can go and what it can execute. More privilege means a larger potential blast radius, but more autonomy is not automatically more valuable.​

The question should be whether the marginal utility of giving an agent more authority justifies the additional exposure. If a recommendation produces most of the business value, allowing autonomous execution may add little utility while significantly increasing the consequence of failure.​

Recent cybersecurity evaluations show why this matters. OpenAI disclosed that models circumvented controls and reached external systems, including Hugging Face. Anthropic has separately reported Claude models gaining unauthorized access to real third-party systems during evaluations. These were not ordinary enterprise deployments, so they should not be treated as forecasts of future claims. They do, however, show how quickly capability, connectivity and permissions can combine once autonomous systems are given room to act.​

That creates a problem for static underwriting. An agent can gain new permissions, connect to additional systems or reveal new behaviors after a policy is bound. The annual policy may not have changed, but the authority budget has.​

An AI policy cannot show whether an agent’s authority is actually bounded. Before deployment, the approval process should connect the business case to the access being granted, the controls being tested and the human review that remains in place for higher-risk actions. This is where responsible AI has to move from principle to practice.​

The value of any framework depends on whether it helps teams document those choices in a way security leaders, boards and insurers can review. For insurers, that evidence may become as important as traditional security controls. A company should be able to show how machine authority was requested, approved, tested, logged and revoked as the agent’s role changes. Without that record, an AI policy says little about the risk created when an authorized system makes the wrong call.​

There are two questions I would urge all leaders to ask: What is the most consequential action our AI can take without human approval? Has that authority expanded since the risk was assessed?​

If the authority budget has expanded, the financial exposure has expanded with it. That is the point cyber insurance will have to absorb. In the agentic AI space, security posture is no longer the whole story. The ability to prove that machine authority is intentional, bounded, observable and reversible may become just as important to insurability as proving that the network itself is secure. That leads me to the main question we should ask: Can you tell me what your AI agent did last week?​

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?