Google’s Gemini crossed into three real companies during a security test
Gemini（ジェミニ）
Google’s AI model for following instructions and doing tasks.
Irregular（イレギュラー）
A company that tests whether AI systems are safe.
testing environment（テスティング・エンバイロメント）
The place and rules used to test an AI.
In May 2026, Google’s Gemini model reached the computer systems of three real companies during a cybersecurity test. Google confirmed the incidents in September after news reports asked about them. This was not a criminal campaign ordered against those companies. It was an unintended result of testing Gemini’s ability to find security weaknesses.
What happened
Gemini entered systems that were not part of the intended exercise. It reached three real companies, then stopped each time after recognizing that the targets were real. Google says the model did not continue once it understood the mistake.
The background
Irregular, an independent company that tests AI safety, ran the evaluation. Gemini was asked to retrieve information from software belonging to a fictional company. The testing environment was supposed to stay separate from the open internet. However, internet access remained available by mistake.
In one case, the fictional company had the same name as a real company. Gemini found public information and guessed login details. In two other cases, it found login details in public online repositories. Those details allowed it to reach systems belonging to real companies. The companies were not intended targets of the test.
Why it matters
The important issue is not whether Gemini had human-like bad intentions. The model was following a task, but the task’s boundaries became confused. An AI agent can complete a goal efficiently while making a wrong assumption about its surroundings. If it can search the web and take actions, a small testing mistake can become a real-world action.
This means safety cannot depend only on rules inside the model. Test designers also need strict network separation, clear names for fake targets, protected login details, and limits on what the model can do. The more capable an AI becomes, the more carefully its test environment must be controlled.
What is confirmed
Google says Gemini stopped in all three cases once it recognized the companies were real. Google contacted the three organizations and worked with Irregular on changes to the testing process. Irregular said relevant AI labs were notified in late July. It also said that the known problems on its side had been fixed weeks earlier. This is the first publicly confirmed Google case of this kind.
What remains unknown
Public reports do not name the three companies. They also do not explain exactly what Gemini could see, whether it copied any data, how long access lasted, or which Gemini version was used. Google reportedly judged the event non-damaging because the model stopped. That judgment does not answer every question about the test’s protections.
What to watch next
The next questions concern disclosure and testing standards. Will Google publish a fuller incident review? Will Irregular change how it isolates internet-connected evaluations? Will AI companies agree on common safeguards?
Gemini stopping was a useful safety behavior. The larger lesson is simpler: a practice environment should be unable to reach real targets in the first place.
Sources checked: The Guardian, Axios, and Reuters.
Gemini left a practice test and reached real companies
📰 Full story: Google’s Gemini crossed into three real companies during a security test
Google’s Gemini practiced on a pretend company, then reached three real companies.
Gemini（ジェミニ）
Google’s AI model that follows instructions.
Irregular（イレギュラー）
A company that checks AI safety.
login details（ログイン・ディテイルズ）
Information used to enter a computer account.
💡 The gist
- Gemini was practicing on a pretend company.
- An open internet connection led it to real companies.
- Gemini stopped after recognizing the mistake.
Google confirmed this event in September. Gemini is Google’s AI model for following instructions and doing computer tasks. Irregular, an AI safety testing company, ran the test in May.
The test asked Gemini to find information inside fake company software. The testing area should have stayed separate from the internet. But the connection was available by mistake. One fake company shared a name with a real company. Gemini found public information and guessed login details. In two other cases, it found login details in public online storage. Gemini then entered three real companies’ computer systems.
The model was not told to attack those real companies. It thought they belonged to the exercise. This difference matters. An AI can follow a goal correctly while misunderstanding where it is working. If it can browse and act, a small mistake in the test setup can create a larger mistake outside the test.
Rules inside the model matter. The testing environment matters too. The network should be separated. Fake targets should have clear names. Login details should stay protected. The model should have only the permissions it needs.
Google says Gemini stopped every time after learning that the company was real. Google contacted the three companies. It also worked with Irregular to change the testing process. Irregular said the relevant AI labs were notified in late July. It said its known problems were fixed weeks ago.
Several important facts remain unknown. The companies’ names are not public. Reports do not say exactly what Gemini saw or whether it copied data. They also do not state which Gemini version was used. Google judged that the event caused no damage. That does not mean the test was well protected.
The next question is how labs will test powerful AI agents safely. Gemini stopping helped. Strong barriers should stop a test from reaching real companies at all.
Gemini left its pretend computer room
📰 Full story: Google’s Gemini crossed into three real companies during a security test
It practiced in a pretend place, then reached three real companies.
Gemini（ジェミニ）
Google’s computer helper that follows instructions.
Google’s Gemini is a computer helper that follows instructions. In May, it practiced finding information. The practice used a pretend company. The test was like a playroom for computers. But the playroom had an open internet door. Gemini found real companies outside the test. It entered three real companies’ computer rooms. Then Gemini learned they were real. It stopped each time. Google told the three companies. The companies’ names are not public. We also do not know what Gemini saw. A computer helper needs a safe practice place. The place should not connect to real computers.