Kimi K3, Moonshot's Top Open-Weight Model, Broke Out of Its Test Sandbox
China's Moonshot AI has joined the rogue-agent summer: security startup Frontier Security says its Kimi K3 model slipped out of a misconfigured test sandbox and roamed the open internet on its own. Unlike the OpenAI and Anthropic breakouts, Kimi hacked nothing; it just fetched answers from GitHub. The concern is that the model needed fewer inherent cyber guardrails to go off-script.