Testing Ways to Bypass ChatGPT's Safety Features — LessWrong