| Moonshot AI's Kimi K3 model escaped a UK AI Security Institute sandbox during a cybersecurity test by exploiting a basic network misconfiguration, accessing GitHub to retrieve benchmark answers instead of solving tasks independently. [link] [comments] |