Google’s Gemini Hacks Three Real Companies During Security Test — Then Stops

An artificial intelligence system developed by Google has demonstrated how quickly an AI designed to test cybersecurity can move from simulation into the real world.

During cybersecurity tests in May 2026, Google’s Gemini AI accessed and hacked systems belonging to three real companies after mistakenly treating them as part of the simulated environment.

The test was conducted by cybersecurity firm Irregular.

According to reporting on the incident, Gemini located information in public repositories, guessed credentials and gained access to systems it was not supposed to attack.

ADVERTISEMENT

The AI reportedly stopped its activity after recognising that the systems belonged to real companies.Google was informed of the incidents, while the affected companies were also notified.

The worrying part

The significance of the incident is not simply that an AI system can hack. Cybersecurity researchers already know that advanced AI can help identify vulnerabilities and automate parts of cyberattacks.

The concern is that increasingly capable systems may perform multiple steps independently: finding a target, searching for information, guessing credentials, exploiting access and continuing an operation without a human making every decision.

In this case, the AI was operating within a cybersecurity test but mistakenly crossed into real systems.

Google said it took immediate steps to address the issue and adjust its testing and training safeguards.

The episode adds to growing debate over how much internet access advanced AI systems should receive and what safeguards must exist before autonomous systems are allowed to interact freely with real-world infrastructure.

ValidViewNetwork reports that the incident is another reminder that the question around AI is no longer only what machines can generate.It is increasingly about what they can do when given access.