Legacy Security Can't Protect Enterprise AI: Build a Way to Stop Your AI Now

"We don't need to spend money on security." Many companies still think this way. We're too small to be a target. Nothing has happened so far, so we're fine.

I believe a cyberattack big enough to go down in history could happen in Japan, too. I'm not trying to scare anyone; I'm saying we should prepare on that assumption. And today, attackers can use AI as well. As attacks get cheaper to run, the range of companies that get targeted widens.

AI now works inside the company

Until recently, AI was a tool where a person asked a question, read the answer, and made the decision. Now AI reads internal data, connects to other tools, and actually carries out work.

Draw a line between outside and inside, and harden the entrance. That has long been how security was done. It wasn't wrong, but when an AI that is already inside reads internal data and exchanges it with outside services, watching the entrance tells you nothing about what is happening inside.

An explainer from the security company Palo Alto Networks makes the point that as AI spreads through enterprises, visibility alone is not enough: companies need real-time control over AI access, usage, data, and spend. They call this an "AI control plane." In my own words, it is "a way for the company to stop or narrow AI's access, data, and spending while it is running." It is a concept, not the name of a specific product.

"Security is expensive" is also a misconception

The flip side of "we don't need to spend on security" is "security costs too much, so we can't touch it." I think this is a misconception too.

Plenty of tools already exist as open-source software, and today you can use AI to build something that fits your own company. Just as AI shows its power when used in cyberattacks, it can be just as powerful for defense. Reading logs for suspicious activity, finding gaps in settings, checking that internal rules are followed: these are all things AI is good at.

A time it helped, and a time we were too late

To be honest, it wasn't by design, but because we had a password manager and MDM (a system for managing company devices centrally), responding when something happened was much easier. The company knew who was using which account and where each device was, and could change or stop things right away. Looking back, that was a small control plane.

What matters in security is being able to change your own setup quickly when something happens at another company. If you don't know what is where, you can't change it even if you want to.

On the other hand, I have also seen failures where being too cautious meant we didn't make it in time. You try to settle on the perfect measures before moving, and the situation changes while you are still deliberating. In the fast-moving world of AI, this is even more true.

The rule of thumb: try it small, with a few people

My rule, in a word, is "try it with a few people first." New security tools and new rules for using AI both start with a handful of people, so you can find and fix what doesn't fit while it is still small. Move quickly and in small steps on decisions you can reverse, and be careful with decisions that are hard to reverse. Choosing where to spend your caution is what matters.

Also, relying on outside services no longer means you are safe. Outside services can be targets too, and the company that understands how its own AI is used is the company itself. That is why I think it is better to use AI in-house as well, to build security measures that fit your company.

Cover three things: access, data, and cost

• Access: decide what each AI may touch, narrow it to what is needed, and make sure you can revoke permissions immediately when you want to stop it
• Data: decide what may be given to AI and what must never leave the company, and reflect that in both rules and systems
• Cost: set a monthly limit, get notified as you approach it, and stop when it is exceeded

If you can see these three and stop them, AI is no longer something you hand everything over to; it becomes a tool under the company's control.

Because you can stop it, you can try it with confidence

Last time, I wrote about making AI adoption an "experiment" by deciding your exit criteria before you start. Exit criteria only work when you have a way to stop. Security is not a brake on using AI; it is what lets you press the accelerator with confidence.

This week, list the AI tools used in your company, and for each one write down what it can access, what data you give it, how much it costs per month, and who would stop it, and how, if you wanted to. The blanks you find are your first step.