Props to OpenAI for publishing this post on some safety and alignment issues observed in internal deployments - there are many counter-incentives to publishing stuff like this, but by making it public we all get better info about safety at the frontier.