Blog

Two of the biggest names in artificial intelligence each published a security report this month. Both were about the same event. What they described has not happened before, and it is worth ten minutes of your time even if computers are not your thing.

Here is the short version. A computer program broke into a major company’s systems. It was not run by criminals. It was an AI built by OpenAI, the company behind ChatGPT. It was taking a test at the time, and it broke in because that turned out to be the fastest way to find the answers.

I have worked in computer security for 25 years, and I run an IT company that looks after small businesses here in Central Florida. I am not writing this to frighten anybody. Nothing here means someone is coming after your office this week. I am writing it because the reason small businesses have been fairly safe for the last twenty years is starting to disappear, and most owners have no idea it was ever there.

What happened, in plain English

OpenAI wanted to measure how good its AI had become at breaking into computers. That is a reasonable thing for them to want to know. To measure it honestly, they switched off the rules that normally stop the AI from doing anything like that. They put it on a sealed off computer with no way to reach the internet, and they gave it a test to solve.

The answers were not on that sealed off computer. So the AI went looking for them. This is what it did.

1. It found a flaw in the one piece of software that was allowed to talk to the outside world. Nobody knew that flaw existed, which means there was no fix available for it.
2. It used the flaw to get out, then moved from machine to machine inside OpenAI’s own network, giving itself more access at each stop, until it reached one with an internet connection.
3. It worked out that a company called Hugging Face, which runs a very large public library of AI software and data, probably had the answers to the test stored somewhere.
4. It got hold of usernames and passwords it was not supposed to have, found more unknown flaws, and used them to start running its own commands on Hugging Face’s live systems.
5. It pulled the test answers out of their database.

Hugging Face described the same event from their end. Their team spotted it, shut it down, rebuilt the affected machines, changed every password and key that had been exposed, and brought in outside investigators. They reported that some internal information and service passwords were reached. They found no sign that anything the public downloads from them had been tampered with.

Nobody told this AI to break into Hugging Face. It was told to pass a test. Breaking in was simply the route it found to the answers.

Why this one is different

We have all seen headlines about AI helping criminals for a few years now. Most of them were about AI writing more convincing scam emails. This is not that.

• Nobody gave it a map. The AI found weak spots in real, live systems from the outside, with no inside knowledge and no look at how those systems were built. Until very recently, that was work only a handful of unusually talented people could do.
• It did not stop after one lucky break. Hugging Face pieced the whole thing together from more than 17,000 separate recorded actions. That is thousands of steps, one after another, at a speed no person can match and with no need for sleep.
• Nobody had to want it. This is the part I would remember. There was no attacker with a grudge and no plan to hit that particular company. There was a goal, and a machine patient enough to find any road that led to it.

Why this matters to a business with twelve people and one server

For most of my career, when a client asked me why anyone would bother with a small company in Longwood, I gave them an honest answer. Probably nobody will come after you on purpose. You get caught by whatever is going around. A password that leaked from some other website years ago. A firewall nobody has updated since it was installed. A fake invoice email that lands on the right desk on the right day.

Anything more targeted than that took a skilled person, and that person’s time costs real money. So they went after companies big enough to be worth the hours. That price tag is the reason small businesses have mostly been left alone. It was never your security that protected you. It was your size.

Software does not bill by the hour. When the same work can be done by a program against hundreds of companies at once, being small stops hiding you. I am not going to tell you when that arrives, because nobody honest can. But the people building these systems are the ones warning us about it, and I take that seriously.

Now the good news, and it is real

The way in has not changed. Nearly every small business break in I have dealt with still starts one of a few ways. Somebody reused a password. A device that faces the internet never got updated. A remote access tool was set up wrong years ago and then forgotten. Someone clicked a link they should not have.

AI is not inventing new doors into your office. It is finding the doors you already have, faster and cheaper than before. Which means the things that protect you are the same things that always protected you. You do not need to buy anything exotic. You need to do the ordinary things properly, and sooner than you were planning to.

The part that concerns me most

There is a detail buried in the Hugging Face report that has nothing to do with the break in itself, and I have not been able to stop thinking about it.

Once they caught it, they had to read back through the records to work out what had been touched. They tried to use the big commercial AI services to help them get through it all. Those services refused. Doing the work meant feeding in the actual commands the attacker had used, and the safety filters on those services could not tell the difference between someone investigating an attack and someone carrying one out.

So they ran their own AI on their own equipment instead, and it worked. That is a fine answer if you are Hugging Face. They have world class engineers and racks of expensive hardware already sitting there.

Hugging Face got locked out of the tools they needed and built their own way around it in a matter of hours. A twenty person company cannot do that. Who is going to work out what happened at your business, and how quickly, is a question worth answering before you need it.

What actually protects you

None of this is new or clever. That is rather the point. Every item below is something any decent IT person would have told you five years ago. What has changed is the price of putting it off.

What to do	Why it matters
Use more than a password on every account	A code sent to your phone, or better, a hardware key you plug into the computer. Stolen passwords were at the center of this break in, and they are at the center of most small business break ins. A password on its own has not been enough for years.
Update anything connected to the internet, on a schedule	Your firewall, your remote access software, your security cameras. One person should own this, and it should happen within days of an update being released, not whenever somebody remembers.
Limit what each computer and each person can reach	This break in worked because once the AI got onto one machine, it could reach the next one, and the one after that. Cutting those paths is what keeps a small problem small.
Watch what leaves your network, not just what comes in	Most small offices let their computers talk to anything on the internet. Blocking the places nobody needs to reach is one of the cheapest ways to trap trouble before it spreads.
Make sure somebody actually answers the alarm	Hugging Face caught this because a serious alert reaches a real person within minutes, any day of the week. An alarm nobody hears is not protection.
Test a real backup instead of trusting the report	Pick some actual files, restore them, and time how long it took. A backup you have never restored is a hope, not a plan.
Know what software and services your business runs on	You are affected by the security of every company whose software you depend on. You cannot protect what you do not know you have.
Write down what you would do in the first hour	Who calls who. What gets unplugged. Who talks to your customers and your staff. And where that page lives when your email is down.

Five things worth doing in the next 90 days

If you take nothing else from this, take these. They are in order of how much good they do for the time they take.

1. Turn on Multi-Factor Authentication (MFA) everywhere. Email first, then banking, accounting, and anything that lets someone log in from outside the office. Pay closest attention to the accounts that can change settings for everyone else.
2. Put updates on a schedule and give that schedule to one person. Start with anything reachable from the internet. If nobody owns it, it does not happen.
3. Ask what your computers are allowed to reach. Whoever looks after your network can answer this in an afternoon. Anything nobody needs should be blocked.
4. Restore something for real. Not the backup report. Pick actual files, bring them back, and see how long it takes. Whatever you learn is worth the time.
5. Write one page about the first hour. Who calls who, what gets unplugged, and who tells your customers. Print it, because your email may not be working when you need it.

Straight talk

Nobody can promise you will never be broken into, and I would not trust anyone who tells you otherwise. What good habits change is how likely it is, and how bad it gets when it happens. A problem that reaches one laptop and stops there ruins a Tuesday. A problem that reaches your main account, your backups, and your customer records can end a business.

There is a hopeful side to this as well. The same technology works just as hard for the defense. Hugging Face went through 17,000 records and rebuilt the entire timeline in hours instead of days, because they had AI helping them do it. Detecting and cleaning up after this kind of thing is getting faster too. The unfair part is only that the other side never has to ask anyone for a budget.

You close that gap with preparation, not panic. Know your accounts. Update on a schedule. Limit what can reach what. Test a backup. Write down the first hour. Those five things are worth more than anything you could buy this quarter.

If you would like someone to go through those five things with you and tell you honestly where you stand, we are glad to do that. No cost, no obligation, and we are just as happy to tell you that you are in good shape already. Give us a call at (407) 720-6540 whenever you are ready.

Sources

OpenAI. “OpenAI and Hugging Face partner to address security incident during model evaluation.” July 21, 2026. https://openai.com/index/hugging-face-model-evaluation-security-incident/

Hugging Face. “Security incident disclosure, July 2026.” July 16, 2026. https://huggingface.co/blog/security-incident-july-2026