For most of the history of cybersecurity, we have had a reasonably clear idea of the problem we were trying to solve. Someone attacks the organization, exploits a vulnerability, steals credentials, compromises information or disrupts a system. Security teams build controls intended to prevent those things from happening and develop response and recovery capabilities for the times when prevention fails. Cyber insurance developed around much the same model. An event occurs, a loss follows, and the policy determines which parts of that loss the insurer has agreed to absorb.

AI is beginning to make that model considerably more complicated because we are moving from technology that primarily provides information to technology that can take action. An AI system that summarizes a document or recommends a course of action creates one kind of risk. An AI agent that can use credentials, interact with systems, communicate externally or execute a business process has been given something more consequential. It has been given some degree of authority to act on behalf of the organization.

Several developments over the past few days show why that distinction matters. They also expose a larger issue that goes well beyond AI. Businesses are becoming increasingly dependent on technologies and external services without always reconsidering what happens when those dependencies behave differently than expected or disappear altogether. That is where the discussion moves beyond cybersecurity and into Business Survivability.

When Software Becomes Part of the Decision

The risks surrounding autonomous AI became much more tangible last week when California Attorney General Rob Bonta issued an investigative subpoena to OpenAI as part of an ongoing investigation into cybersecurity incidents and risks associated with AI models. According to the California Department of Justice, the investigation includes an incident involving Hugging Face and questions surrounding whether AI systems can be manipulated into performing malicious activity.

The significance of the investigation isn't limited to the technical details of what an AI agent may or may not have done. The more important issue is responsibility. The California Attorney General's announcement specifically emphasized that companies developing frontier AI models have responsibilities to prevent their technology from perpetrating or enabling cyberattacks.

Reuters reported that the California investigation comes as the Federal Trade Commission is conducting its own industry-wide examination of potential risks associated with AI systems. Other state attorneys general have also sought information concerning the Hugging Face incident. What is beginning to emerge is a regulatory expectation that increasing autonomy doesn't somehow remove accountability from the organizations developing or deploying the technology.

This matters because businesses are rapidly moving beyond using AI as an assistant. They are beginning to give AI agents access to systems and information that allow them to participate directly in business operations. Once that happens, AI governance becomes much more than a discussion about acceptable use or data privacy. The organization has made a decision about how much authority a piece of technology should have.

We already understand this concept when people are involved. Organizations decide who can approve payments, change production environments, access sensitive information or make commitments on behalf of the company. Controls develop around those decisions because the consequences of exercising authority incorrectly can be significant. AI shouldn't somehow exist outside that same logic simply because the actor is software.

The challenge is that the delegation of authority to AI may happen much more gradually than it does with people. A team adopts a tool because it makes a particular task easier. Someone connects it to another system. Additional permissions are granted as its usefulness becomes apparent. A vendor introduces AI functionality into a product the business already uses. Over time, something that began as a productivity tool can become part of the way the business actually operates without anyone going back and reconsidering the risk created by the accumulation of those decisions.

That is where decision debt begins to appear. The problem isn't necessarily that any individual decision was wrong. The problem is that a series of individually reasonable decisions can eventually create an operating environment that nobody intentionally designed and that few people fully understand.

Insurance Will Have to Follow the Authority

This is also where AI begins to create difficult questions for insurance. A recent Insurance Business article examined what happens when an AI agent goes beyond its instructions and causes a loss. The article highlights the difficulty of determining responsibility among the organization using the technology, the developer that created it and the contractual relationships governing its use.

For the business deploying the AI, the uncomfortable reality is that responsibility may remain much closer to home than many organizations expect. The ability to recover a loss from an AI provider may depend on contractual protections and the ability to demonstrate what actually caused the system to behave incorrectly. Insurance Business also points to the growing importance of traceability because establishing what happened may require records showing the instructions the system received and the actions it subsequently took.

That creates an interesting extension of a problem we are already seeing in cyber insurance. Underwriters have become increasingly interested in evidence showing that security controls aren't simply documented but are actually operating as represented. AI takes that evidence problem into an entirely new area. If an AI agent makes a decision that results in a loss, the business may need to demonstrate what the agent was authorized to do and whether the controls intended to limit its authority were working.

The question for an insurer will eventually become more complicated than whether the organization uses AI. Insurers will need to understand how AI participates in the operation of the business and what consequences can follow from the authority it has been given. That becomes especially difficult when the organization itself doesn't have a complete inventory of where AI is being used or how deeply it has become embedded in important processes.

This gets very close to the question I have been exploring through Can We Insure This? Insurance works best when we understand the risk we are asking someone else to absorb. If the organization cannot clearly explain how an AI agent is being used, what it can access and what happens when it makes a mistake, then the problem begins before anyone opens the insurance policy.

Accountability Is Already Becoming Unclear

Another recent Insurance Business report illustrates how unsettled accountability already is. Research involving more than 2,500 CIOs and CTOs found that 52% said the CIO was being held accountable when an AI agent made an error.

That may be convenient organizationally, but it doesn't necessarily make sense from a governance perspective. AI happens to be technology, but the decisions being delegated to it may belong to finance, operations, customer service, human resources or another business function. Treating every AI failure as a technology failure can obscure who actually decided to use the system and who benefits from the decisions it makes.

The insurance consequences become equally complicated. The same Insurance Business report points out that AI exposures may intersect with directors and officers insurance, cyber coverage, technology errors and omissions, and professional liability. A single event may not fit comfortably within the traditional boundaries organizations have become accustomed to using when thinking about insurance.

That doesn't mean every AI deployment requires an entirely new insurance product. It means businesses need to understand how AI changes their existing exposures before assuming that an existing policy will respond in the way management expects. That understanding begins with governance rather than insurance.

The Other Side of the Problem Is Whether the Business Can Keep Operating

While the insurance industry and regulators are beginning to work through the consequences of AI autonomy, another set of recent findings points to a more familiar problem. Organizations still struggle to maintain operations when cyber events become serious.

TechRadar Pro recently reported research showing that 47% of CISOs who had experienced a cyber incident said the event resulted in an operational shutdown. More troubling is that 73% of CISOs surveyed said they weren't confident they could effectively manage a major cyber incident if one occurred tomorrow.

Those findings should change the way we think about cyber resilience. Much of the traditional discussion has focused on whether systems can be restored and how quickly that restoration can occur. Those are important questions, but they don't necessarily tell us whether the business can continue operating while recovery is taking place.

PwC's newly released 2027 Global Digital Trust Insights reinforces that concern. PwC surveyed 3,934 business and technology leaders across 71 countries and found that only 39% of security, risk and operations leaders had fully formalized and integrated operational continuity plans specifically addressing cyber threats. The same research found that attacks against AI systems were the cyber threat organizations felt least prepared to address.

Taken together, those findings suggest that organizations are increasing their dependence on AI while still struggling with the more basic problem of maintaining operations through technology disruption. That is a dangerous combination because the operating environment is becoming more complicated at the same time the organization's ability to function without parts of that environment remains uncertain.

Recovery and Survivability Are Not the Same Thing

Recovery has always been an important part of cybersecurity, but I think we sometimes allow the concept of recovery to narrow the problem too much. A recovery plan can tell us how quickly a particular application should be restored. It doesn't necessarily tell us what happens to the company while that application is unavailable.

The business still has customers who expect service and employees who need to work. Financial obligations continue even while systems are down. Contracts don't disappear because the network is unavailable. Management still needs reliable information to make decisions, often at exactly the moment when normal sources of information have become unreliable.

This is why I increasingly think about cyber resilience in terms of the minimum viable business rather than simply the recovery of technology. The objective during a serious disruption isn't necessarily to restore every capability immediately. The first priority is preserving enough operating capability for the organization to remain viable while recovery occurs.

TechRadar's reporting describes organizations beginning to think in similar terms by defining the minimum capabilities required to continue operating during a crisis. That approach moves the conversation from a list of systems toward an understanding of the business functions those systems support.

The difference matters. A recovery plan may tell the technology team which application needs to come back first. A Business Survivability approach should also explain what the organization does if that application doesn't come back when expected.

The Business Doesn't Control All of Its Dependencies

That problem becomes more difficult because many of the capabilities businesses depend on are no longer inside the organization.

A recent Insurance Business article examines the growing concern among cyber insurers around third-party dependencies. Spencer Timmel, Head of Cyber Insurance at Safety National, describes vendor risk as both a cybersecurity and business resilience issue because compromise of a provider can affect many customers simultaneously. The article also cites Verizon's 2026 Data Breach Investigations Report, which found third-party involvement in 48% of breaches in its dataset.

The breach statistic matters, but the operational dependency behind it matters more. Businesses have spent years moving important capabilities to cloud providers, software platforms and specialized service providers because doing so often makes economic and operational sense. What sometimes gets lost is that outsourcing the capability doesn't transfer all of the consequences associated with losing it.

A company can have excellent internal security controls and still experience a serious operational disruption because a provider fails. The company's endpoint protection may work exactly as intended. Its employees may use MFA and its backups may be secure. Its cyber insurance may be properly structured. None of that guarantees that the business can continue operating when a critical external dependency becomes unavailable.

This is one reason cybersecurity maturity and Business Survivability shouldn't be treated as interchangeable concepts. Good cybersecurity contributes to survivability, but survivability requires understanding a much broader operating environment.

AI Is Making Those Dependencies Harder to See

AI makes the dependency problem even more difficult because it can become embedded inside services the organization already uses. A company may know that it depends on a particular software provider without knowing which AI models or services that provider depends upon to deliver its product. The organization may not know whether those AI capabilities have become essential to the service or what happens if they suddenly become unavailable.

The same problem exists internally. Employees may introduce AI into workflows without intending to create a new dependency. A task that was initially accelerated by AI can eventually become difficult to perform without it. As more workflows develop that way, AI stops being a separate technology initiative and becomes part of normal operations.

At that point, describing the problem as "AI risk" becomes less useful. It is operational risk created partly by AI, and it needs to be understood in the same context as any other capability the business depends upon.

That is also where the weekend's stories begin to converge. The regulatory scrutiny surrounding autonomous AI agents is fundamentally about what happens when technology is given authority. The insurance discussion is about who bears the financial consequences when that authority produces an unexpected outcome. The continuity research is showing how difficult it can be for organizations to operate when important technology becomes unavailable. The vendor-risk discussion reminds us that much of the technology businesses depend upon isn't even under their direct control.

These aren't separate problems. They are different expressions of the same change in how businesses operate.

Business Survivability Starts Before Something Goes Wrong

When an incident happens, executives understandably want to know whether insurance will cover the loss. That question matters, but by the time it is being asked, most of the decisions that determined the organization's exposure have already been made.

The more important work happens earlier, when the business decides which technologies it will depend on and how much authority it will give them. It happens when someone determines how much access an AI agent needs to perform its job. It happens when a critical business process is moved to an external provider and when management decides what evidence is sufficient to believe that a control or recovery capability actually works.

Those decisions accumulate over time. Some create efficiencies and some reduce costs. Others allow the organization to grow in ways that wouldn't otherwise be possible. They also create dependencies, and those dependencies eventually determine what happens when something fails.

Insurance remains an important part of managing the financial consequences. Cybersecurity remains essential to reducing the likelihood and impact of attacks. Recovery planning remains necessary for restoring technology.

Business Survivability sits across all of them because the ultimate issue isn't whether a particular control worked or whether a particular loss was insured. It is whether the organization can continue functioning when one of the assumptions it built its operations around turns out to be wrong.

AI is making that question harder because software is no longer simply supporting business decisions. Increasingly, it is participating in them. At the same time, businesses are becoming more dependent on technology and providers they don't completely control.

That makes the question behind Can We Insure This? more important, but it also makes it insufficient on its own. Before we can determine whether a risk can be insured, we need to understand what the business has become dependent upon, who or what has been given authority to act, and what remains operational when one of those dependencies fails.

Insurance can help a company absorb the financial consequences of an event. It cannot make an organization survivable after the fact.