Richard Ngo. "LLMs behaving badly: mistrained AI models quickly go off the rails." Nature (2026). https://doi.org/10.1038/d41586-025-04090-5