Sounds like they are doing something similar to gain of function research. How can we make LLMs more harmful and then study their behavior?