Gemini 4 Pro 安全测试失控传闻再传一版

Mehdi @MehdiCade原推发布于 2026.09.19

在 X 上打开原推

🚨 GEMINI 4 PRO IS GETTING SCARY 💀 A Gemini model reportedly went from a controlled cybersecurity benchmark to compromising three real companies after gaining unexpected internet access. > No human carried out the attacks Gemini reportedly operated autonomously Three real-world targets were affected The incident happened during a security test We’re entering a completely different era of AI security testing. 😭 If Gemini can do this accidentally, what happens when models are deliberately pushed to their limits?