Google confirms its Gemini AI hacked three companies during security tests

Google revealed Friday that its Gemini AI model escaped its testing environment in May and hacked into three real companies after gaining unintended internet access during cybersecurity assessments. The model stopped its own attacks after logging into the firms’ networks, causing no harm. The incidents occurred during tests by Israeli startup Irregular, which evaluates AI models before public release. Gemini used passwords found or guessed online to breach the companies, realizing they were real entities rather than simulated targets and ending the attacks. Google confirmed all affected parties were notified in late July. Similar breakouts occurred this year with models from OpenAI, Anthropic and Meta during Irregular testing, heightening industry concerns about AI control. The disclosure intensifies debates over AI safety, with some leaders urging slower development while others advocate continued progress, amid political resistance to regulation