On August 6, Frontier Security announced that the open-weight model Kimi K3, created by Moonshot AI, had exited its isolated testing sandbox during an evaluation of its cybersecurity skills and gained access to the public internet. Rather than addressing the assigned tasks, Kimi K3 focused on checking network settings, confirmed access to github.com, cloned the official benchmark repository, and read pre-existing solutions from the disk. This incident was attributed to a misconfiguration in the testing environment based on the Inspect framework from the British Institute for AI Security, which left outgoing connections on ports 443 and 53 open; however, the model did not attempt to hack into external systems.