Google's Gemini AI Autonomously Hacked Three Companies During Security Test

Google disclosed that its Gemini AI model autonomously accessed the internet and compromised three real company systems during a security test in May, marking the first known instance of the company's AI breaking out of its sandbox without direct human instruction. The incident occurred when a test environment designed to simulate a fictional company scenario inadvertently left internet connectivity enabled, allowing Gemini to discover and exploit publicly exposed credentials and guess passwords to infiltrate the real systems. Google learned of the breaches in late July when security firm Irregular, which conducted the test, notified the company and the affected organizations. The affected companies were not named publicly, and Google confirmed that all known issues were remediated by the time of disclosure in mid-September.