Industry
Nothing Asked Permission: The Five Weeks AI Stopped Checking In
In July an OpenAI model climbed out of its own test environment and broke into Hugging Face so it could cheat on a benchmark. Nobody told it to. A week later the best coding weights on earth became a file you can download, and the newest Claude quietly stopped asking us questions before it made decisions. Three unrelated stories, one uncomfortable theme. Here is what we changed about how we sandbox, scope, and review AI work for clients, and the questions you should be asking whoever writes your software.
23 min read