On a scale from one to ten, how in control is the average large company of the AI running inside it? At VERGE: Comprehensive Security, a panel from Lovable, Kyndryl and Ethira answered three, three and a half, and five. Here’s how to raise that number.

Today, anyone in your organisation can ship an app, connect an agent or onboard a new AI vendor before lunch. That’s great for innovation. It’s slightly terrifying for the board.

The Build Fast. Build Safe. panel brought together three levels of AI safety. Alejandra Arreola Ruiz works on AI safety and product policy at Lovable, the platform that lets millions of people build software. Joakim Lundberg leads security and resilience in the Nordics at Kyndryl and has sat in a bank’s CISO chair. Lucas de Araújo started out as an ethical hacker, worked in security at Klarna and is now co-founder and CTO at Ethira.

The panel opened with a quick round. On a scale from one to ten, how in control is the average large company of the AI running inside it? The answers were three, three and a half, and five.

Three questions most companies can’t answer

What AI is running here? Who owns it? Can we prove it’s behaving?

Most companies, even those with a CISO and a compliance team, struggle with all three. Lucas explained why. Security teams have historically been understaffed, and AI makes tracking changes harder still. Knowing which AI you have only matters if you also know what it can access: which databases, which data sets, which data flows. Getting that to a meaningful level is a real puzzle.

And AI rarely hides out of bad intent. It usually appears where people are simply trying to do their job, connecting company data to faster, nicer tools than the ones they were given. The answer is for organisations to provide those tools themselves.

“We built it over the weekend. Can we launch Monday?”

Joakim’s answer to the weekend build is three questions:

1. Do you truly understand the risk?
2. Are you authorised to accept that risk?
3. Is it legal?

If the answer to all three is yes, so be it. The point isn’t to block, but to make sure someone with the mandate has actually accepted the risk.

The bigger problem is time. The window between a vulnerability becoming public and being exploited has shrunk from weeks to days. A launch-Monday culture leaves almost no room for a traditional security review.

Layers, not patches

Joakim’s conclusion was that the traditional approach of finding and patching every vulnerability is critically flawed. AI will keep finding new ones faster than anyone can patch them.

The answer is old wisdom applied with new urgency: defence in depth and zero trust. A bank may not be able to stop someone getting into an account, but it can stop the money leaving. Limits, time delays and monitoring are all layers. Build enough of them, and one will catch what the others miss.

Prevent. Detect. Respond.

Your suppliers are the new front door

Asked whether he would go in through the suppliers if he were breaking into a company today, Lucas didn’t hesitate. Yes.

Suppliers are usually outside an organisation’s own security processes, and most third-party security is still handled through yearly questionnaires. Lucas wants to see that replaced by real-time visibility into the security of the companies you depend on. If your payments provider has a gap, you should know right away, not at next year’s assessment.

Secure by design, not by reminder

Joakim explained secure by design with a story about his kids. You can ask them to remember to lock the door. Maybe they will, maybe they won’t. Or you can install a lock that locks itself and opens with a fingerprint. Then nobody has to remember, and a lost key doesn’t matter.

That’s what good governance looks like. Lucas made the same point about AI governance: it’s not a new word for a compliance department that slows everyone down. It means giving developers frameworks where the secure and compliant way is also the easiest way.

Alejandra described how this works at Lovable’s scale. You start by understanding how users actually behave on the platform and which risks are emerging. You build on established principles, such as international law and human dignity, and the real work is translating those principles into daily operations. And you create spaces where people are invited to break things, internally and with outside experts, because no company can find its own blind spots alone.

Proof, not promises

Regulators and boards increasingly expect leaders to prove control, not just claim it. Joakim described what proof looks like in practice:

• A foundation. Control of your data and documented processes for assessing risk.
• A process per use case. Describe the use, assess whether you should do it, define the controls, then let it loose and monitor it.
• Evidence. All of this creates logs and documents that serve as proof.
• Independent assurance. Someone from outside who asks to see the results, the process and how you test your controls.

Lucas pointed to the next step: moving beyond one-time attestations towards showing, continuously, that you work securely in practice.

Where to never compromise

Asked where a company should accept risk to move fast, and where it should never compromise, the panel landed close to where the afternoon began.

Alejandra: if you’re building a car, you never compromise on steering and brakes. Know what is critical, and keep doing risk assessments.

Lucas: never transfer the risk to your users or other parties.

Joakim: agree on what you will and won’t do before the contract, not after something goes wrong. Doing it right from the start is often faster than fixing it later.

The question nobody is asking yet

Joakim left the room with a question to take to the afterwork. One day, many companies will have almost as many AI agents as employees. Each agent holds context, memory and half-finished work. If something goes wrong, how do you recover that state? Has anyone thought about that before rolling out agents?

What to do on Monday

• Ask your team the three questions: what AI runs here, who owns it, and can we prove it behaves?
• Don’t ask whether you have an AI policy. Ask how it’s put into practice.
• Name your three most critical suppliers and check how you would know if they had a security gap today.
• Agree who has the mandate to accept risk on a fast launch.
• Add another layer of defence wherever you rely on a single control today.

About the panel

• Alejandra Arreola Ruiz works on AI safety and product policy at Lovable.
• Joakim Lundberg leads security and resilience in the Nordics at Kyndryl.
• Lucas de Araújo is co-founder and CTO at Ethira, a former ethical hacker who previously worked in security at Klarna.

This article is part of VERGE Pulse, a series of insights from VERGE: Comprehensive Security at Epicenter Stockholm on 19 October 2026. VERGE is Epicenter’s flagship event series, bringing founders, leaders, investors and experts into one room to explore the questions that will shape the next decade.

Want to talk about resilience in your own organisation, or shape the next VERGE with us? Get in touch with the Epicenter team.