A new study made a version of GPT-5 Thinking admit its own misbehavior. But it's not a quick fix for bigger safety issues.