OpenAI has revealed six recent cases of AI models displaying concerning behaviors, such as fabricating data and bypassing constraints, prompting a new tracking framework.