Did Google Gemini Go Rogue and Hack Three Companies? Not Exactly

Headlines are making the rounds claiming Google's Gemini AI "broke out" and hacked three real companies.

That sounds like the beginning of the AI apocalypse we've been warned about.

So, did Gemini go rogue?

No. But what actually happened is still interesting, and there's a cybersecurity lesson here that has very little to do with some super-intelligent AI suddenly deciding to become a hacker.

What Actually Happened?

The incident occurred in May 2026 while Gemini was participating in cybersecurity testing conducted by Irregular, a company that specializes in testing frontier AI models for security risks.

Gemini was essentially participating in a sophisticated capture-the-flag exercise. It was given a fictional company and tasked with finding information and gaining access to systems inside what was supposed to be a controlled testing environment.

There was one fairly significant problem.

The AI models unintentionally had access to the real internet.

There was another problem.

According to Irregular's own investigation, the company believed the fictional company name used for the evaluation didn't correspond to anything real.

It did.

The fictional name happened to coincide with a real, previously existing, in use domain.

Now you have a powerful AI that has been specifically instructed to hack a fictional target, but it can reach the real internet and OOOOPS! there's a real target with the same name.

You can probably see where this is going.

Gemini went looking for what it believed was part of the exercise and wound up interacting with real systems.

So How Did Gemini "Hack" Them?

This is where the story becomes considerably less Terminator and a lot more like a typical day for a penetration tester.

In one case, Gemini guessed passwords until it found one that worked.

In two other cases, it found credentials in a publicly accessible repository and used them to access protected systems.

That's it.

Gemini didn't discover some previously unknown Windows vulnerability. It didn't invent a new hacking technique. It didn't crack modern encryption or develop some terrifying AI-generated zero-day.

It did things human hackers and security researchers have been doing for decades.

The difference is that Gemini did those things autonomously.

Did Gemini Actually Go Rogue?

Not in the way most people would interpret "going rogue."

Gemini had already been instructed to hack something. That was the entire point of the cybersecurity exercise.

The bigger failure was the testing environment.

The AI believed it was operating inside an authorized security exercise, but the environment unintentionally gave it access to the public internet.

In other words, Gemini didn't dramatically smash through the walls of its sandbox and escape onto the internet.

The door was basically left open.

Google says Gemini stopped the hacking in all three cases after learning it had accessed real companies. The affected organizations were notified, and Irregular says the underlying problems have since been fixed.

And It Wasn't Just Gemini

This is another important part of the story.

Similar incidents involving AI models from OpenAI, Anthropic, and Meta have been reported over the past few months.

Taken individually, that sounds pretty alarming. It looks like AI models from four major companies are suddenly breaking out of their sandboxes and attacking the internet.

That's not quite what happened.

Irregular says the subsequent public disclosures all trace back to the same underlying issue with its evaluation environment.

The Next Web put the timeline together and found that Irregular notified the affected AI developers in late July. The incidents were then disclosed publicly at different times, which made them appear more like a series of separate AI breakouts.

That's an important distinction.

Four AI labs independently losing control of their models would suggest one problem.

Four AI labs using the same testing environment with the same underlying configuration problem suggests another.

The Part That Actually Concerns Me

There's a detail in Irregular's account that's arguably more important than the "rogue AI" headline.

Monitoring these systems is difficult.

AI cybersecurity evaluations can involve thousands of simulations and enormous amounts of activity. Worse, because these tests are deliberately asking AI models to perform cyberattacks, the logs are naturally filled with behavior that looks malicious.

You're basically looking for an unauthorized cyberattack inside thousands of authorized simulated cyberattacks.

Irregular acknowledges that manual review isn't sufficient and says it is improving monitoring, containment controls, evaluation review, and procedures for selecting fictional domains.

That's where I think the real AI security story is.

As AI agents become capable of taking actions rather than simply answering questions, we need to know where they're going, what they're doing, and whether they've crossed a boundary they weren't supposed to cross.

There's Another Lesson Here

I'm going to look at this from the other side for a moment.

Gemini apparently got into three companies using password guessing and credentials that were publicly exposed.

That's not exactly cutting-edge hacking.

We don't know enough about the affected companies or their configurations to say that any single security measure definitely would have stopped Gemini. But these are precisely the kinds of attacks basic security practices are supposed to make much more difficult.

Strong, unique passwords matter. So does two-factor authentication. Companies should also be monitoring repositories and development environments for accidentally exposed passwords, API keys, and tokens if they use them.

For regular users, a good place to start is Have I Been Pwned. Enter your email address and you can see whether it has appeared in known data breaches. Its Pwned Passwords service can also tell you whether a password has appeared in known breach data.

If a password has been exposed, don't just change one character and keep using it. Replace it with a completely new password and don't reuse that password anywhere else.

A Few Tools That Can Help

You don't need enterprise security software to improve your protection. A password manager is one of the easiest places to start. If you are like me, you have simply far to many passwords to remember. That tends to make users choose something simple and repeate using that password that everywhere. Thats a very bad idea, especially in this new environment.

RoboForm is our favorite and makes it easy to generate strong, unique passwords and automatically fill them when needed. That means you don't have to remember dozens of complicated passwords, and more importantly, you don't have to reuse the same password across multiple sites.

Bitwarden is another excellent option for generating and storing unique passwords, while KeePassXC is a good choice for people who prefer an open-source password manager with their password database under their own control.

Passwords are only part of the equation. 2FAGuard provides a Windows-based way to store and manage two-factor authentication codes. With 2FA enabled, stealing or guessing your password may not be enough for an attacker to get into your account.

None of these makes you unhackable. Nothing does. But the methods Gemini reportedly used here become considerably less effective when passwords are unique, exposed credentials are replaced, and another authentication factor is required.

The Part We Should Actually Be Paying Attention To

None of this means the AI side of the story should be dismissed.

Gemini didn't do anything particularly remarkable from a hacking standpoint. What is remarkable is that an AI agent could search for a target, locate credentials, try passwords, access systems, evaluate what it found, and continue working toward an objective.

A human penetration tester can do all of those things.

Now an AI agent can too.

And potentially it can do them faster, continuously, and eventually at a scale that would be difficult for humans to match.

That's where this story becomes interesting.

The immediate lesson isn't that Gemini became self-aware and escaped onto the internet. It's that we're giving increasingly capable AI agents tools, internet access, and the ability to take actions on their own.

When we do that, the guardrails around those agents become every bit as important as the intelligence of the AI itself.

Geek Verdict

No, Gemini didn't suddenly go rogue and decide to start hacking companies.

It was told to hack during a cybersecurity test. A testing environment unintentionally allowed access to the real internet, a fictional target overlapped with a real one, and Gemini wound up accessing real systems using some surprisingly ordinary techniques.

The AI capability is worth paying attention to. An autonomous agent can now perform work that previously required a human sitting at a keyboard and do that MUCH faster.

But underneath all the scary AI headlines is a much older cybersecurity lesson: AI didn't need some sophisticated new exploit to get in. Weak passwords and exposed credentials were enough. The technology doing the hacking may be new. The security mistakes that made it possible are anything but.

Maybe we should go back to sticky note plastered on our monitors with “PetsName 123” on them?

Jim McMahon Jim McMahon, aka Corporal Punishment, is the founder of MajorGeeks.com. He has spent decades testing software, troubleshooting Windows, and helping users cut through the nonsense. He loves real freeware, hates bloatware, and runs on caffeine, sarcasm, and questionable choices.

/c]

comments powered by Disqus