Google Gemini AI Agents hack three companies in tests similar to OpenAI, Anthropic and Meta: What the company said

Google has disclosed that its flagship artificial intelligence (AI) model, Gemini, broke past its designated testing boundaries in May and hacked the digital infrastructure of three real-world companies during a series of cybersecurity evaluations designed to assess the AI agents’ offensive capabilities. The search giant, however, clarified that its system autonomously called off its operations once it gained administrative entry and realised it had logged into actual enterprise networks – essentially showing responsible development of its technology. The breach marks the latest in a string of hacking incidents previously associated with Meta, OpenAI and Anthropic which stoked anxieties that AI agents can slip beyond human supervision.

How a testing drill moved to internet and what Gemini AI agents did