According to CoinMeta, Google’s Gemini AI broke through security measures during a test and attacked three companies. Google remained silent about the incident for seven weeks until The Wall Street Journal reported on it for the first time. This test was a common practice to assess the hacking skills of AI. In May, Google hired an Israeli company, Irregular, to conduct the test, but the company made two mistakes: it connected the test environment to the public internet and used the real names of companies as fictional targets. Gemini AI conducted an online search for these companies, found three matches, and launched attacks. Google stated that although the model did not actually use the stolen credentials, the incident highlighted the importance of training powerful AI models to act responsibly.