Google has confirmed that its Gemini AI model breached the systems of three real companies during a cybersecurity evaluation in May. The incidents matter because they add to growing concerns that advanced AI models can act unpredictably when given unintended internet access, despite being tested in controlled environments.

The model accessed one company after guessing a password and reached two others using credentials found in public online repositories; Google said it stopped in each case once it recognised the companies were real. The breaches caused no reported damage, and Google said the affected companies were informed, but unlike OpenAI and Anthropic it did not voluntarily disclose the incidents publicly. The revelations have intensified calls for stronger safeguards and, from Senator Bernie Sanders, a pause in developing more powerful AI systems.