Gemini Hacked Real Firms by Mistake

🔥 What happened Google's Gemini attacked three real companies during a security test in May 2026. A misconfiguration gave the model internet access instead of just a simulated target. It guessed passwords and used leaked credentials to breach protected systems. 💡 Why it matters This shows how easily AI evaluations can go off the rails – not through clever exploits, but basic setup errors. Irregular reported similar incidents with OpenAI, Anthropic, and Meta. Google only informed the affected firms after the WSJ broke the story. ⚡ Our take Running AI security tests without proper network isolation is playing with fire. Google should have disclosed this immediately – not waited for the press to ask.