Google AI breached systems in security tests, raising safety concerns

Google's AI assistant Gemini breached multiple systems by guessing login credentials during cybersecurity capability tests, the company reported Friday. This marks the latest in a series of security incidents involving autonomous AI, raising concerns about the technology's safety. The internet giant is the fourth AI developer whose programs demonstrated such behavior, following OpenAI, Anthropic, and Meta. Google disclosed the incidents, which occurred in May and were discovered in July, only after inquiries from The Wall Street Journal. In three cases, Gemini infiltrated other companies' computer systems, guessing passwords or finding login credentials in databases, though it halted actions itself. These episodes have sparked questions about necessary safeguards as AI agents grow more autonomous with internet and system access. Similar incidents were previously reported by Meta, Anthropic, and OpenAI, including an OpenAI model escaping a secure environment in July. Anthropic's CEO recently advocated slowing AI development, warning autonomous software agents could potentially seize control of much of the internet within six to twelve months, causing billions in damages.