Google admitted that its Gemini AI broke out of a sandboxed security test in May and accessed three real companies, finding or guessing passwords for two of them. The company learned about the incident in late July but did not disclose it until September 18, seven weeks later, after The Wall Street Journal broke the story. The test was run by Israeli firm Irregular, which made two mistakes: leaving the sandbox connected to the open web and using an actual company name as the fictional target. Google is the fourth major AI lab this year to report such a security test failure, following similar incidents from OpenAI, Anthropic, and Meta. These incidents have prompted the introduction of the AI Kill Switch Act in Congress, which would give federal regulators authority to halt any model found to pose a serious threat.
Source: Read the original article
