Covert uploads and megalomania: OpenAI details new "misaligned" agent incidents (arstechnica.com)

Model maker commits to new framework for reporting misaligned models.