The New Economics of Cybersecurity

AI is making some parts of security dramatically cheaper. That does not mean cybersecurity is becoming cheaper.

I have spent enough time around vulnerability reviews, remediation discussions and production changes to know that finding the problem is rarely the end of the story. Quite often, it is the beginning of the expensive part.

A vulnerability may be technically obvious. What happens next is much less straightforward. Is it actually exploitable in our environment? Which application or business service depends on it? Who owns the system? Can we patch it today? What needs to be tested first? What happens if the change fails? Can we roll it back? And, somewhere in that chain, who makes the decision when security urgency collides with operational risk?

AI is starting to change the economics of this process, but not evenly.

In the first two articles of this series, I looked at what frontier AI changes for cyber risk and why remediation is becoming such an important constraint. My conclusion was not that models such as Claude Mythos or OpenAI Daybreak suddenly create an entirely new category of cyber risk. They change something more practical: the speed, scale and accessibility with which existing weaknesses can be discovered, analysed and potentially exploited.

They also make some defensive activities considerably faster. What interests me now is what happens to the economics when one part of the security lifecycle accelerates much faster than the rest. AI can make a finding cheap. It does not automatically make the decision, the fix or the production change cheap. And that distinction matters.

The Cost Curve Is Bending Unevenly

Some of the clearest evidence can be found in something much less exotic than autonomous exploitation: phishing. A 2026 field experiment presented at USENIX Security found that AI-generated personalised phishing emails could be produced for roughly $0.03 per message while achieving a 54% click-through rate, comparable with skilled human red-teamers in the experiment. A click is not a compromise, so I would not stretch that result further than the study allows. But the economic implication is difficult to ignore.

Personalisation used to consume human time. Someone had to research the target, understand the context and write the message. If much of that work can now be replicated for a few cents, an attacker does not necessarily need a dramatically more capable technique to change the economics. Being able to apply an existing technique across thousands of targets can be enough.

Time, expertise and human attention have always constrained cyber operations. AI is weakening those constraints for certain repetitive or research-intensive tasks.

That does not mean we have reached the era of the autonomous hacker.

The evidence there is still much less convincing. Some offensive AI systems have achieved success rates above 85% in favourable known-vulnerability benchmark scenarios, only to fall to roughly 13% under more realistic CVE-Bench conditions and lower still against difficult targets. That gap is important.

The near-term change is not that expert attackers suddenly become unnecessary. It is that some of the work surrounding an attack—reconnaissance, personalisation, analysis, iteration—can be performed at much lower marginal cost. For defenders, the picture is more complicated.

Defenders Get the Technology, But Not Quite the Same Economics

Security teams can use the same class of technology for vulnerability analysis, source-code review, alert triage, investigations, detection engineering, exploitability assessment and remediation support. There is good reason to be optimistic about some of this.

IBM's 2026 breach research, covering 602 organisations, found that organisations reporting extensive use of AI and automation in security operations also reported average breach costs almost $2 million lower than non-adopters. I would still be careful with the interpretation. Correlation is not causation, and mature organisations may simply be better at both adopting automation and containing incidents. But it is credible evidence that AI can improve defensive economics when it is embedded into a mature security operation.

The part I find more interesting is what happens when AI moves from identifying a problem to fixing it. Academic research on AI-generated software patches shows why this is not straightforward. A patch can pass a test suite and still be larger, more complex or more difficult to reason about than an equivalent human change. Other research has shown that increasingly autonomous patch-generation workflows can introduce new vulnerabilities while fixing the original problem.

From an enterprise perspective, this makes perfect sense. Generating candidate code and authorising a production change are two very different things. AI may produce a plausible fix in seconds. Someone still needs to understand the affected system, review the change, run regression testing, consider dependencies, schedule the deployment and decide whether the remaining uncertainty is acceptable.

The coding effort may fall substantially while much of the assurance effort remains. That is where I think the economics become more useful than the headline capability.

Where the Money Actually Goes

We often talk about vulnerability management as if it has one cost.

In practice, I see at least four.

The first layer is straightforward. Discovery is becoming cheaper.

The second is where enterprise context enters the picture.

A finding does not arrive with a complete understanding of the organisation attached to it. We still need to know whether the issue is reachable, whether an attacker can realistically exploit it, which service depends on the asset, whether compensating controls exist and how much business impact a remediation action could create.

Most of that information lives across different teams and systems. Architecture knows one part. Security another. Infrastructure, application owners and business teams know others. Sometimes nobody has the complete picture.

AI can help assemble that context. It cannot magically create context that the organisation itself does not have.

Then comes remediation.

Someone has to make the change. That can involve development work, vendor coordination, testing, application owners, maintenance windows, deployment, validation and rollback preparation.

And finally there is the cost we sometimes understate in security discussions: the risk introduced by the fix itself.

A badly executed security change can interrupt a critical business service just as effectively as an attack.

So when I think about security value, I increasingly think about it in fairly simple terms:

How much verified reduction in exposure did we achieve, and what did it cost the organisation to achieve it?

That is a more useful economic question than how many vulnerabilities a tool can find.

What Happens When Findings Become Cheap?

This is where AI creates a slightly uncomfortable possibility. Better discovery can make the overall security operation more expensive. Wiz's Red Agent is an interesting recent example. During its first month of public preview, the autonomous red-teaming capability reportedly surfaced more than 17,000 findings across roughly 1,000 environments. That is impressive scale.

But imagine the same effect inside a large enterprise without changing anything else in the operating model. Every credible finding starts another process. Someone needs to validate it. Someone needs to determine whether it matters. An owner needs to be identified. Engineering work may need to be scheduled. Testing needs to happen. A change may require approval. Production has to be modified and, finally, somebody needs to verify that the original exposure has actually disappeared.

Anyone who has managed vulnerability remediation at scale will recognise the problem. The bottleneck is rarely that the organisation desperately needs another list of vulnerabilities. The difficult part is deciding which problems matter enough to interrupt other work and then getting the organisation to act on them safely.

If AI makes the first step ten times cheaper and generates ten times as much output, but the rest of the organisation remains unchanged, we have not eliminated the cost. We have moved it. And in a poorly designed operating model we may simply have built a much more sophisticated backlog.

That is why I would be cautious about measuring the value of AI security products primarily through findings generated. Discovery has economic value only when the organisation has enough capacity downstream to convert the information into reduced exposure.

AI Also Has a Meter Running

There is another aspect of the economics that I think deserves more attention. Traditional enterprise security software has generally been relatively predictable commercially. We became accustomed to pricing per endpoint, per user, per asset, by data volume or through a negotiated enterprise agreement.

Agentic AI introduces something different: reasoning itself has a marginal cost.

Every model call consumes compute. Larger context windows cost more. An agent that decides it needs six additional iterations through a codebase consumes more than one that stops after two. At small scale, this can appear trivial.

Across millions of lines of code, continuous cloud analysis, autonomous validation and large numbers of applications, it becomes an architecture and operating-cost question. Model choice matters. Context management matters. Caching matters. Batch processing matters. Routing simple tasks to cheaper models while reserving frontier models for harder reasoning matters.

Current public frontier-model pricing can vary by roughly an order of magnitude and, in some cases, considerably more between model tiers. This means the architecture surrounding the AI becomes part of the economics of the security service itself.

I originally approached this part of the research wondering whether enterprises could eventually face what I thought of as a subsidy cliff: organisations becoming dependent on artificially inexpensive inference and later discovering that the real price was much higher. I could not find strong enough evidence to support that argument, so I would not make it.

The more practical concern is less dramatic anyway: consumption can be variable, model prices can change, and autonomous workflows make future operating costs less predictable than the deterministic tools security teams are accustomed to.

For a CISO or CIO trying to build a multi-year business case, that matters.

Remediation Has Its Own Economics

There is another misconception that AI may force us to revisit: that patching faster is always the correct answer. Security teams understandably focus on the cost of waiting. Leave an exploitable vulnerability open and the opportunity for compromise remains. If active exploitation exists, every additional day matters.

But there is also a cost to acting.

Anyone who has been involved in an urgent production change knows what that means. Applications can fail. Dependencies can behave differently than expected. Emergency changes can create incidents. The pressure to move quickly sometimes encourages us to compress the same validation steps that protect the business from a badly implemented fix.

Published estimates of downtime vary enormously, so I would not attach too much significance to one universal number. But the broader reality is clear: for a large enterprise, serious outages can cost hundreds of thousands of dollars per hour and much more under the wrong circumstances.

So “patch everything faster” is not really a risk strategy.

The real question is whether we can reduce security exposure fast enough without increasing operational risk beyond what the business can tolerate.

That requires more than a CVSS score.

It requires exploitability, reachability, business criticality, compensating controls, maintenance requirements, confidence in testing and a clear understanding of what happens if the change goes wrong. Security does not protect the organisation by winning its own metric while the business becomes unavailable.

This is also why I believe investment in automated testing, deployment safety and rollback capability increasingly belongs in the cybersecurity conversation. Those may look like engineering or resilience investments, but they determine how quickly security exposure can actually be removed.

Some AI Economics Is Really Governance Economics

One result from IBM's 2026 research stood out to me for another reason. It reported that 92% of organisations breached through their own AI systems had not implemented basic access controls beforehand. The easy headline is that AI introduces new security risk.

The more useful observation is that weak fundamentals become more expensive when technology allows them to operate at greater speed and scale. I think the same principle applies much more broadly. Unclear application ownership. Incomplete asset inventories. Technical debt. Weak change processes. Poor data governance. Ambiguous risk appetite. None of these problems started with generative AI.

AI simply exposes the consequences faster.

This is why I am sceptical when AI security discussions jump immediately to another technology purchase. In many organisations, the return on improving ownership, architecture knowledge, testing or remediation workflows may be greater than the return on generating another layer of findings.

AI does not make fundamentals less important.

It makes the price of weak fundamentals more visible.

What Should the Next Euro Buy?

This brings me to the question security leaders eventually have to answer.

How much should we spend?

It is a reasonable question, but I increasingly think it is the wrong place to start.

Global cybersecurity expenditure keeps increasing, while Gartner's board research indicates declining confidence among directors that existing cybersecurity investments are delivering the level of protection they expect. That does not prove organisations are spending too much. And board confidence is not an objective measure of security performance. But it should make us question whether budget size is a useful proxy for protection.

A €20 million security programme is not automatically twice as good as a €10 million programme.

The question I would rather answer is:

What additional protection are we buying with the next euro?

Can that investment reduce the exposure window for a known exploited vulnerability? Can it shorten the recovery time of a critical service? Can it remove an unsupported platform that continually creates exceptions? Can it improve testing enough that critical patches reach production safely in two days instead of ten? Can it give an application a clear owner where today nobody is accountable? Can it eliminate hours of manual validation that are preventing skilled people from working on the risks that actually matter? Those questions are harder to answer than “How much are we spending?”

Questions we are asking ourselves every day.

They are also much closer to the outcome the business is paying us to deliver. And this is where AI should influence investment priorities. If discovery becomes progressively cheaper, the next euro may not belong in another discovery capability. It may create more value in prioritisation, automated testing, engineering capacity, deployment automation, rollback, asset ownership, technical-debt reduction or recovery. The answer will be different for every organisation because the constraint will be different. That is the point.

Invest where the constraint actually is.

The Metric I Would Watch

AI will make parts of cybersecurity dramatically cheaper. It will also generate more work, create new forms of consumption cost and expose organisational weaknesses that no model can solve on its own.

So I do not think the organisations that benefit most from AI in cybersecurity will necessarily be the ones deploying the greatest number of agents or generating the greatest number of findings. I would watch something less exciting.

How efficiently does the organisation convert security investment into verified reduction of meaningful exposure?

That brings the economics back to the outcome. Finding ten thousand vulnerabilities is not the outcome. Closing ten thousand tickets is not the outcome. Spending more than last year is certainly not the outcome. The outcome is that an exposure capable of becoming a business event was identified, understood and reduced—and that we managed to do it without creating another business event in the process.

AI can make many individual steps in that chain extraordinarily fast. The challenge for security leaders is making the organisation around those steps fast enough to benefit. That, more than the price of the next model or the number of findings it can generate, is where I believe the new economics of cybersecurity will be decided.