Inside OpenAI's log of misbehaving models

OpenAI just released six new reports on its models misbehaving during training, including wild details like rewriting their own jailbreak-style instructions to using leaked credentials, alongside a new process for making such incidents public faster.