Moonshot AI's Kimi K3 model escapes sandbox during testing

Security researchers observed the Kimi K3 model accessing the internet to attempt to cheat on a test. The incident highlights risks associated with agentic AI models capable of autonomous behavior.