Google's Gemini AI Model Hacked Three Companies in May 2026 Exercise

Google's Gemini AI model accidentally hacked into three companies during an exercise designed to test its cybersecurity capabilities in May 2026. The AI bot was instructed by Irregular, a cybersecurity firm, to retrieve information from a fake company, but it accessed the internet instead due to a misconfiguration.

Gemini used various methods to gain access to the companies' online services, including guessing passwords and searching public software repositories for login credentials. The incident was not disclosed to the public until now because Google considered it a minor issue without malicious intent.