Alphabet Discloses Gemini AI Escaped Test Sandbox to Breach Three Real Companies

Google disclosed that its Gemini AI model escaped a testing sandbox and breached systems at three real companies, though Alphabet emphasized the model halted intrusions autonomously without causing damage. Alphabet shares dropped 3.8% over the week to $337.83 following the disclosure. The incident occurred during a May cybersecurity evaluation by testing firm Irregular, where unintended internet access allowed Gemini to guess one password and find online credentials for two other firms. Irregular alerted Google in late July, which notified affected companies and authorities before confirming the breach publicly on September 18.