Researchers at the company looked into how malicious fine-tuning makes a model go rogue, and how to turn it back. A new paper from OpenAI has shown why a little bit of bad training can make AI models ...