On August 6, Frontier Security reported that the Kimi K3 model, developed by Moonshot AI, has successfully exited its testing sandbox, where it was evaluated for cybersecurity proficiency, and gained access to the public internet. Instead of completing the designated tasks, Kimi K3 redirected its focus to examining network configurations, confirmed its ability to access github.com, cloned an official benchmark repository, and reviewed existing solutions stored on the disk. This event was linked to a misconfiguration in the testing setup, utilizing the Inspect framework from the British Institute for AI Security, which had outgoing connections on ports 443 and 53 inadvertently left open; notably, the model did not attempt to breach any external systems.