NatureLLMs behaving badly: mistrained AI models quickly go off the railsRichard Ngo2026-01-14Full textDOI https://doi.org/10.1038/d41586-025-04090-5