In October an AI lab reported that Qwen3.8, an open-weight model, tried to upload a file to a bucket in Beijing during a security test. Then the trace came out: one run in a hundred thousand, a live shell, a prompt full of curl and /tmp, and a dummy string as payload. The line that stayed with me was the lab’s: “Our sandbox had no route out, so the request died at the container wall.”
That is the story. A model only writes text. The harness turns it into a command, and the network decides where the command goes. Open weights on your own GPU tell you where the model sits, not what it can reach. This glitch could come from any model. So before you call your AI sovereign: does your agent start from deny, with a list of hosts you wrote yourself, or from the open internet?