Frontier AI security: faster discovery does not mean managed risk 

Frontier AI is changing the speed of vulnerability discovery. It’s getting a lot of attention from security leaders, executives, and boards. Rightly so, sort of…the main question is not whether AI can find more vulnerabilities. It is whether organizations can understand, prioritize, and act on what AI finds.

Most companies are not suffering from a lack of findings. They already have vulnerability data, alert volume, security tool outputs, compliance gaps, vendor reports, and remediation backlogs. Adding AI to that environment may improve discovery, but it does not automatically improve risk management.

A vulnerability only becomes meaningful when it is connected to business context. What system is affected? Is it exposed? Is it exploitable in this environment? Does it support revenue, operations, client delivery, regulated data, or executive priorities? Who owns the fix? What gets deprioritized if this gets funded first?

Without those answers, faster discovery can create faster confusion.

Quick answer

• Frontier AI can speed up vulnerability discovery, but discovery is not the same as risk management.

• A vulnerability only matters when it is tied to asset criticality, exposure, exploitability, ownership, and business impact.

• Most companies should not build their own AI vulnerability tools.

• The better move is to strengthen asset visibility, exposure management, attack path mapping, secure SDLC, and remediation ownership.

• AI increases the value of a mature cyber program. It does not replace one.

• There is no substitute for the fundamentals: asset management, access controls, reduced complexity, security monitoring, and a complete cyber risk program.

What frontier AI security gets right

David Jones’ Cybersecurity Dive article on frontier AI and vulnerability discovery points to a Palo Alto Networks update from Lee Klarich covering Claude Mythos, Claude Opus 4.7, and GPT-5.5-Cyber. The articles highlighted a real shift: AI models are getting better at finding security vulnerabilities quickly, and vendors are already using them to accelerate research and remediation.

There was a massive increase in vulnerabilities discovered: it included 26 CVEs across more than 130 products, compared with its typical monthly volume of fewer than five. None were being exploited in the wild at the time of disclosure, but the volume shows why security teams are paying attention.

AI materially increases the speed and scale of discovery; this means attackers will move faster and defenders will need to adjust.

But AI is not magic. Palo Alto’s own writing makes that clear. Klarich notes that useful results require more than asking a model to find bugs. The process needs context, scanning harnesses, guardrails, threat intelligence, and multiple models working together.

That is the part leaders need to pay attention to. Board members and CEOs shouldn’t run to their CISOs and CIOs and start demanding this.

AI can accelerate discovery. It still needs a program around it.

Discovery is not exploitation 

Marketing loves a dramatic leap from “AI found a vulnerability” to “AI will break everything.” But that leap skips operational reality.

Finding a vulnerability is not the same as exploiting it. Exploitation depends on reachability, configuration, access, compensating controls, monitoring, segmentation, identity posture, and how the affected system sits inside the larger business environment. Simply put as a business leader “Do I care about this?”

A critical vulnerability on an isolated, low-value system is not the same as a medium vulnerability on an internet-facing system that supports revenue, production, or client data. Severity is useful, but severity is not priority by itself. 

If I’m being honest, this is where many programs struggle already, without AI. They have scanners. They have dashboards. They have lists. What they often lack is the connective tissue between technical findings and business decisions.

That connective tissue is the program.

Asset visibility tells you what exists. Business impact analysis tells you what matters. Exposure management tells you what is reachable. Attack path mapping shows how weaknesses can compound. Governance tells you who can accept, reduce, transfer, or fund the risk.

Without that structure, AI does not produce clarity. It produces more data.

The Mythos lesson is still fundamentals 

Ross Haleliuk made an important point in his recent posts about Mythos and the AI security conversation. The right response is not panic or tool-chasing. It is a return to the fundamentals that have mattered for years: know what you have, reduce unnecessary exposure, enforce least privilege, monitor well, strengthen access controls, and build security programs around people and process.

 “What works today against Mythos is what worked 5 years ago before Mythos - reducing attack paths, controlling lateral movement, enforcing least privilege, validating the environment, etc. As you can see, everything has changed, but also very little has changed." - Ross Haleliuk.

That may sound less exciting than frontier AI, but it is where serious security programs are won. The most important stuff is usually the most boring stuff.

Most organizations already know they have issues. They do not need a new model to tell them there is too much access, too little visibility, weak segmentation, unmanaged assets, slow remediation, or unclear ownership. They need leadership alignment and a practical operating model that turns those issues into funded, prioritized work.

This is why the answer to frontier AI is not simply more AI. It is better security management.

Most companies should not build their own AI vulnerability tools 

I loved C. Kelly Bissell’s LinkedIn post about his home-made vulnerability tool “Ava” is brilliant because it shows both sides of the conversation. He built a lightweight endpoint malware detection tool in two days and used it to explore whether AI changes the build-versus-buy calculation for security tools.

That is a fair question. AI is lowering the cost of software creation, and experienced practitioners can now prototype ideas faster than ever.

But prototype speed is not the same as production responsibility.

Most organizations should not build and maintain their own AI vulnerability products. Serious security tooling is not just code. It is telemetry, validation, quality assurance, threat intelligence, integrations, support, legal review, update cycles, disclosure workflows, product security, documentation, and long-term maintenance.

Most CISOs already have more responsibility than capacity. Asking them to own a homegrown AI vulnerability tool adds a new layer of accountability. If the internal tool misses an issue that an established vendor would have caught, the organization still owns the outcome. The CISO still has to explain the decision. The board still has to understand why the business accepted that risk.

For most companies, the better path is not becoming a security software vendor. The better path is using strong vendors, demanding better outcomes, and building the internal program maturity needed to act on the information those tools provide.

Where leaders should focus now

Frontier AI helps highlight something most of us in the industry already knew. There’s a lot we are not seeing, and if I can be cliché’, “you don’ t know, what you don’t know.” 

This should push leaders to improve the security program, not throw more data on the already large bonfire security leaders are dealing with. 

The first priority is asset visibility. If the business does not know what it owns, where it lives, who owns it, and how it supports operations, every security decision starts with fog.

The second priority is business criticality. Not every system deserves the same funding, urgency, or control depth. Companies need to understand which assets support revenue, client trust, operational continuity, regulated obligations, and recovery priorities.

The third priority is exposure management. Leaders need to know what is internet-facing, misconfigured, over-permissioned, unmanaged, or reachable through vendor access, identity paths, cloud connections, or stale infrastructure.

The fourth priority is attack path mapping. Vulnerabilities rarely matter in isolation. Risk often comes from how weaknesses combine across identity, network, application, data, and operational systems.

The fifth priority is secure SDLC and DevSecOps. If AI can find flaws faster, companies need to prevent more flaws before they reach production. That means secure design, code review, dependency management, testing, secrets management, and clear ownership inside engineering workflows.

The final priority is security monitoring & remediation capacity. You have to find the bad things happening in order to react to them, and better discovery without the ability to fix things creates frustration.

Build the program before chasing the findings

Frontier AI will make vulnerability discovery faster.

That is useful.

It will help vendors improve products, researchers move faster, and defenders find issues earlier.

But faster discovery does not replace cyber risk management.

The companies that benefit most from frontier AI will not be the ones chasing every new finding. They will be the ones that understand their environment, know what matters to the business, and have a cybersecurity program capable of turning findings into decisions.

AI can find the vulnerability.

Leadership still has to manage the risk.

If your organization is finding more risk than it can prioritize, the issue may not be detection.

It may be program design.

If you want to build a cybersecurity program focused on real risk, visibility, exposure, and business priority, start the conversation.

If you already have a program and want to pressure test whether it is built around what matters most, start the conversation.

Because the real work is not chasing every vulnerability.

It's building the program that knows which ones can actually hurt the business.

Ready to build or pressure test your cybersecurity program?