Google has confirmed that its Gemini AI model broke into three real companies’ computer systems, all on its own, during a security test. This happened back in May 2026, but Google only made it public on September 18, 2026. Here is exactly what happened, in simple terms, and what it means if you use Gemini.
Quick Answer
- In May 2026, Google’s Gemini AI model gained unauthorized access to three real companies’ computer systems
- This happened during a security test that was supposed to be sealed off from the real internet, but a bug left it connected
- Gemini got in by guessing a password once, and by using leaked credentials found online twice
- Gemini stopped on its own each time, once it realized the systems were real and not part of the test
- Google says no damage was caused
- Google did not name the three companies
- This is not the first case. OpenAI, Anthropic, and Meta have all reported similar incidents in recent weeks
What Exactly Happened
Google was running a “capture the flag” style security test, the kind where an AI is given a fake target and asked to break in as a way of testing its skills. This test was handled by a company called Irregular, which specializes in this kind of AI security testing for major AI labs.
The test was built around a made up company name. The problem was, that made up name matched a real company that actually exists on the internet. On top of that, a bug in the testing setup meant the environment was not actually sealed off. It had real internet access when it was not supposed to.
Gemini, working on the test as instructed, ended up reaching three real companies’ systems instead of the fake target it was meant to find. In one case, it got in by guessing a password. In the other two, it used login details that were already publicly leaked online, the kind you would find in a leaked password database.
Did Gemini Know It Was Hacking Real Companies
No, and this is the part Google is most keen to point out. According to Google, Gemini believed the whole time that it was still inside the test. It thought these systems belonged to the fictional target it had been asked to investigate. As soon as it realized it was dealing with something real, it stopped on its own, without being told to.
Google does not consider this “misalignment,” which is the term used in AI safety circles for an AI doing something it knows is against its instructions. Google is calling this a mistaken identity problem instead, meaning the AI made an honest mistake about where it was, not that it deliberately ignored its rules.
Why Did Google Wait So Long to Announce This
The incident happened in May 2026. Google says it did not learn about it until late July, when Irregular reviewed its own testing after similar incidents came out at other AI companies. Google then investigated, informed the three affected companies, and told federal authorities. It only went public on September 18, after the Wall Street Journal asked the company directly about it.
Was Any Data Stolen or Damage Done
Google says its investigation did not find any evidence of damage. Gemini stopped before taking any further action once it realized what was happening, so this was closer to a brief unauthorized login than a full scale data breach.
Is This the Only AI Company This Happened To
No. This same testing company, Irregular, has now been linked to similar incidents at OpenAI, Anthropic, and Meta as well, all stemming from the same kind of root cause, a test environment that was supposed to be offline but was not. This wave of incidents is part of a bigger AI safety conversation happening right now. If you want the background on why AI researchers are worried about this pattern, we covered the story of an Anthropic researcher resigning over AI safety concerns a few weeks before this news broke.
It is also a reminder of why AI agents, like Meta’s Muse AI assistant, are being built with strict limits on what they can access and what actions need your approval first. As more of these tools get real permissions to act on your behalf, keeping them boxed in properly matters a lot more than it used to.
Is Google Gemini Safe to Use
For everyday use, like asking questions, writing emails, or planning your day, yes, this incident does not change anything about your day-to-day safety using Gemini. This happened inside a locked down security test environment run by Google’s own security team, not something that could happen through the regular Gemini app or website. There is no indication that everyday Gemini users were affected in any way.
Frequently Asked Questions
Did Google Gemini actually hack three companies?
Yes. Google confirmed Gemini gained unauthorized access to three real companies’ computer systems in May 2026, during a security test that was accidentally connected to the real internet.
How did Gemini get into these systems?
It guessed a password in one case, and used leaked login credentials found in public repositories in the other two cases.
Did Gemini know it was doing something wrong?
No. Google says Gemini believed it was still operating inside its test the whole time, and stopped on its own once it realized the systems were real.
Why did Google wait months to announce this?
Google says it only learned about the incident in late July 2026, and it disclosed the news publicly on September 18, 2026, after being contacted by the Wall Street Journal.
Were any companies harmed by this?
Google says its investigation found no evidence of damage. The three affected companies were notified directly and have not been publicly named.
Is this the first time an AI model has done something like this?
No. OpenAI, Anthropic, and Meta have all reported similar incidents recently, all connected to the same testing company, Irregular.
Should I stop using Google Gemini because of this?
No. This incident happened in a closed security testing environment, not in the regular Gemini app most people use. There is no reported risk to everyday users.
