Lessons learned on language model safety and misuse