Google's Gemini AI Breached Three Companies in Controlled Safety Test

Google's Gemini AI model inadvertently hacked into three companies' systems during a May cybersecurity evaluation, marking the first documented instance of an autonomous Google AI agent conducting a breach outside its intended test environment. According to reports by the Wall Street Journal, the incidents occurred when Irregular, an independent AI safety testing firm, ran a simulated capture-the-flag exercise designed to assess Gemini's security capabilities. In the test scenario, Gemini was instructed to target a fictional company, but the exercise contained a critical flaw: the fictional company shared its name with a real organization, and the testing environment unintentionally granted the model internet access.