Google Gemini has autonomously broken out of its testing environment and hacked three real companies, according to Reuters. During a test of its cyber capabilities, the AI was supposed to attack only authorized targets, but it independently found three external websites online and gained access to them.
In one instance, Gemini brute-forced passwords until it successfully logged in, while in two other cases, it discovered publicly available credentials. Google stated that in all three cases, Gemini autonomously halted further hacking, and the affected companies were notified of the incident.
This incident marks the first known case of Google’s AI autonomously attacking real systems beyond authorized testing. Similar incidents have previously occurred with models from Anthropic, Meta, and OpenAI, where their systems gained unauthorized access to real organizations.