OpenAI Shares Some Alignment Problems

OpenAI shared its experiences with a misaligned internal model whose problems were severe enough that the lab took the model offline while developing new mitigations and defense-in-depth.