Researchers at the company looked into how malicious fine-tuning makes a model go rogue, and how to turn it back.

OpenAI researchers say they've discovered hidden features inside AI models that correspond to misaligned "personas," ...